6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "ci/ciObjArray.hpp"
29 #include "asm/register.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/c2/barrierSetC2.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/intrinsicnode.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/machnode.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/parse.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subtypenode.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #include "utilities/bitMap.inline.hpp"
51 #include "utilities/powerOfTwo.hpp"
52 #include "utilities/growableArray.hpp"
53
54 //----------------------------GraphKit-----------------------------------------
55 // Main utility constructor.
56 GraphKit::GraphKit(JVMState* jvms)
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn()),
60 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
61 {
62 _exceptions = jvms->map()->next_exception();
63 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
64 set_jvms(jvms);
65 }
66
67 // Private constructor for parser.
68 GraphKit::GraphKit()
69 : Phase(Phase::Parser),
70 _env(C->env()),
71 _gvn(*C->initial_gvn()),
72 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
73 {
74 _exceptions = nullptr;
75 set_map(nullptr);
76 debug_only(_sp = -99);
77 debug_only(set_bci(-99));
78 }
79
80
81
82 //---------------------------clean_stack---------------------------------------
83 // Clear away rubbish from the stack area of the JVM state.
84 // This destroys any arguments that may be waiting on the stack.
842 if (PrintMiscellaneous && (Verbose || WizardMode)) {
843 tty->print_cr("Zombie local %d: ", local);
844 jvms->dump();
845 }
846 return false;
847 }
848 }
849 }
850 return true;
851 }
852
853 #endif //ASSERT
854
855 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
856 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
857 ciMethod* cur_method = jvms->method();
858 int cur_bci = jvms->bci();
859 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
860 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
861 return Interpreter::bytecode_should_reexecute(code) ||
862 (is_anewarray && code == Bytecodes::_multianewarray);
863 // Reexecute _multianewarray bytecode which was replaced with
864 // sequence of [a]newarray. See Parse::do_multianewarray().
865 //
866 // Note: interpreter should not have it set since this optimization
867 // is limited by dimensions and guarded by flag so in some cases
868 // multianewarray() runtime calls will be generated and
869 // the bytecode should not be reexecutes (stack will not be reset).
870 } else {
871 return false;
872 }
873 }
874
875 // Helper function for adding JVMState and debug information to node
876 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
877 // Add the safepoint edges to the call (or other safepoint).
878
879 // Make sure dead locals are set to top. This
880 // should help register allocation time and cut down on the size
881 // of the deoptimization information.
882 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
933 }
934
935 // Presize the call:
936 DEBUG_ONLY(uint non_debug_edges = call->req());
937 call->add_req_batch(top(), youngest_jvms->debug_depth());
938 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
939
940 // Set up edges so that the call looks like this:
941 // Call [state:] ctl io mem fptr retadr
942 // [parms:] parm0 ... parmN
943 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
944 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
945 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
946 // Note that caller debug info precedes callee debug info.
947
948 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
949 uint debug_ptr = call->req();
950
951 // Loop over the map input edges associated with jvms, add them
952 // to the call node, & reset all offsets to match call node array.
953 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
954 uint debug_end = debug_ptr;
955 uint debug_start = debug_ptr - in_jvms->debug_size();
956 debug_ptr = debug_start; // back up the ptr
957
958 uint p = debug_start; // walks forward in [debug_start, debug_end)
959 uint j, k, l;
960 SafePointNode* in_map = in_jvms->map();
961 out_jvms->set_map(call);
962
963 if (can_prune_locals) {
964 assert(in_jvms->method() == out_jvms->method(), "sanity");
965 // If the current throw can reach an exception handler in this JVMS,
966 // then we must keep everything live that can reach that handler.
967 // As a quick and dirty approximation, we look for any handlers at all.
968 if (in_jvms->method()->has_exception_handlers()) {
969 can_prune_locals = false;
970 }
971 }
972
973 // Add the Locals
974 k = in_jvms->locoff();
975 l = in_jvms->loc_size();
976 out_jvms->set_locoff(p);
977 if (!can_prune_locals) {
978 for (j = 0; j < l; j++)
979 call->set_req(p++, in_map->in(k+j));
980 } else {
981 p += l; // already set to top above by add_req_batch
982 }
983
984 // Add the Expression Stack
985 k = in_jvms->stkoff();
986 l = in_jvms->sp();
987 out_jvms->set_stkoff(p);
988 if (!can_prune_locals) {
989 for (j = 0; j < l; j++)
990 call->set_req(p++, in_map->in(k+j));
991 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
992 // Divide stack into {S0,...,S1}, where S0 is set to top.
993 uint s1 = stack_slots_not_pruned;
994 stack_slots_not_pruned = 0; // for next iteration
995 if (s1 > l) s1 = l;
996 uint s0 = l - s1;
997 p += s0; // skip the tops preinstalled by add_req_batch
998 for (j = s0; j < l; j++)
999 call->set_req(p++, in_map->in(k+j));
1000 } else {
1001 p += l; // already set to top above by add_req_batch
1002 }
1003
1004 // Add the Monitors
1005 k = in_jvms->monoff();
1006 l = in_jvms->mon_size();
1007 out_jvms->set_monoff(p);
1008 for (j = 0; j < l; j++)
1009 call->set_req(p++, in_map->in(k+j));
1010
1011 // Copy any scalar object fields.
1012 k = in_jvms->scloff();
1013 l = in_jvms->scl_size();
1014 out_jvms->set_scloff(p);
1015 for (j = 0; j < l; j++)
1016 call->set_req(p++, in_map->in(k+j));
1017
1018 // Finish the new jvms.
1019 out_jvms->set_endoff(p);
1020
1021 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1022 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1023 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1024 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1025 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1026 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1027
1028 // Update the two tail pointers in parallel.
1029 out_jvms = out_jvms->caller();
1030 in_jvms = in_jvms->caller();
1031 }
1032
1033 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1034
1035 // Test the correctness of JVMState::debug_xxx accessors:
1036 assert(call->jvms()->debug_start() == non_debug_edges, "");
1037 assert(call->jvms()->debug_end() == call->req(), "");
1038 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1039 }
1040
1041 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1042 Bytecodes::Code code = java_bc();
1043 if (code == Bytecodes::_wide) {
1044 code = method()->java_code_at_bci(bci() + 1);
1045 }
1046
1047 if (code != Bytecodes::_illegal) {
1048 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1184 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1185 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1186 return _gvn.transform( new AndLNode(conv, mask) );
1187 }
1188
1189 Node* GraphKit::ConvL2I(Node* offset) {
1190 // short-circuit a common case
1191 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1192 if (offset_con != (jlong)Type::OffsetBot) {
1193 return intcon((int) offset_con);
1194 }
1195 return _gvn.transform( new ConvL2INode(offset));
1196 }
1197
1198 //-------------------------load_object_klass-----------------------------------
1199 Node* GraphKit::load_object_klass(Node* obj) {
1200 // Special-case a fresh allocation to avoid building nodes:
1201 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1202 if (akls != nullptr) return akls;
1203 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1204 return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1205 }
1206
1207 //-------------------------load_array_length-----------------------------------
1208 Node* GraphKit::load_array_length(Node* array) {
1209 // Special-case a fresh allocation to avoid building nodes:
1210 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1211 Node *alen;
1212 if (alloc == nullptr) {
1213 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1214 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1215 } else {
1216 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1217 }
1218 return alen;
1219 }
1220
1221 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1222 const TypeOopPtr* oop_type,
1223 bool replace_length_in_map) {
1224 Node* length = alloc->Ideal_length();
1233 replace_in_map(length, ccast);
1234 }
1235 return ccast;
1236 }
1237 }
1238 return length;
1239 }
1240
1241 //------------------------------do_null_check----------------------------------
1242 // Helper function to do a null pointer check. Returned value is
1243 // the incoming address with null casted away. You are allowed to use the
1244 // not-null value only if you are control dependent on the test.
1245 #ifndef PRODUCT
1246 extern uint explicit_null_checks_inserted,
1247 explicit_null_checks_elided;
1248 #endif
1249 Node* GraphKit::null_check_common(Node* value, BasicType type,
1250 // optional arguments for variations:
1251 bool assert_null,
1252 Node* *null_control,
1253 bool speculative) {
1254 assert(!assert_null || null_control == nullptr, "not both at once");
1255 if (stopped()) return top();
1256 NOT_PRODUCT(explicit_null_checks_inserted++);
1257
1258 // Construct null check
1259 Node *chk = nullptr;
1260 switch(type) {
1261 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1262 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1263 case T_ARRAY : // fall through
1264 type = T_OBJECT; // simplify further tests
1265 case T_OBJECT : {
1266 const Type *t = _gvn.type( value );
1267
1268 const TypeOopPtr* tp = t->isa_oopptr();
1269 if (tp != nullptr && !tp->is_loaded()
1270 // Only for do_null_check, not any of its siblings:
1271 && !assert_null && null_control == nullptr) {
1272 // Usually, any field access or invocation on an unloaded oop type
1273 // will simply fail to link, since the statically linked class is
1274 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1275 // the static class is loaded but the sharper oop type is not.
1276 // Rather than checking for this obscure case in lots of places,
1277 // we simply observe that a null check on an unloaded class
1341 }
1342 Node *oldcontrol = control();
1343 set_control(cfg);
1344 Node *res = cast_not_null(value);
1345 set_control(oldcontrol);
1346 NOT_PRODUCT(explicit_null_checks_elided++);
1347 return res;
1348 }
1349 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1350 if (cfg == nullptr) break; // Quit at region nodes
1351 depth++;
1352 }
1353 }
1354
1355 //-----------
1356 // Branch to failure if null
1357 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1358 Deoptimization::DeoptReason reason;
1359 if (assert_null) {
1360 reason = Deoptimization::reason_null_assert(speculative);
1361 } else if (type == T_OBJECT) {
1362 reason = Deoptimization::reason_null_check(speculative);
1363 } else {
1364 reason = Deoptimization::Reason_div0_check;
1365 }
1366 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1367 // ciMethodData::has_trap_at will return a conservative -1 if any
1368 // must-be-null assertion has failed. This could cause performance
1369 // problems for a method after its first do_null_assert failure.
1370 // Consider using 'Reason_class_check' instead?
1371
1372 // To cause an implicit null check, we set the not-null probability
1373 // to the maximum (PROB_MAX). For an explicit check the probability
1374 // is set to a smaller value.
1375 if (null_control != nullptr || too_many_traps(reason)) {
1376 // probability is less likely
1377 ok_prob = PROB_LIKELY_MAG(3);
1378 } else if (!assert_null &&
1379 (ImplicitNullCheckThreshold > 0) &&
1380 method() != nullptr &&
1381 (method()->method_data()->trap_count(reason)
1415 }
1416
1417 if (assert_null) {
1418 // Cast obj to null on this path.
1419 replace_in_map(value, zerocon(type));
1420 return zerocon(type);
1421 }
1422
1423 // Cast obj to not-null on this path, if there is no null_control.
1424 // (If there is a null_control, a non-null value may come back to haunt us.)
1425 if (type == T_OBJECT) {
1426 Node* cast = cast_not_null(value, false);
1427 if (null_control == nullptr || (*null_control) == top())
1428 replace_in_map(value, cast);
1429 value = cast;
1430 }
1431
1432 return value;
1433 }
1434
1435
1436 //------------------------------cast_not_null----------------------------------
1437 // Cast obj to not-null on this path
1438 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1439 const Type *t = _gvn.type(obj);
1440 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1441 // Object is already not-null?
1442 if( t == t_not_null ) return obj;
1443
1444 Node* cast = new CastPPNode(control(), obj,t_not_null);
1445 cast = _gvn.transform( cast );
1446
1447 // Scan for instances of 'obj' in the current JVM mapping.
1448 // These instances are known to be not-null after the test.
1449 if (do_replace_in_map)
1450 replace_in_map(obj, cast);
1451
1452 return cast; // Return casted value
1453 }
1454
1455 // Sometimes in intrinsics, we implicitly know an object is not null
1456 // (there's no actual null check) so we can cast it to not null. In
1457 // the course of optimizations, the input to the cast can become null.
1458 // In that case that data path will die and we need the control path
1541 }
1542
1543 //=============================================================================
1544 //
1545 // parser factory methods for MemNodes
1546 //
1547 // These are layered on top of the factory methods in LoadNode and StoreNode,
1548 // and integrate with the parser's memory state and _gvn engine.
1549 //
1550
1551 // factory methods in "int adr_idx"
1552 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1553 int adr_idx,
1554 MemNode::MemOrd mo,
1555 LoadNode::ControlDependency control_dependency,
1556 bool require_atomic_access,
1557 bool unaligned,
1558 bool mismatched,
1559 bool unsafe,
1560 uint8_t barrier_data) {
1561 assert(adr_idx == C->get_alias_index(_gvn.type(adr)->isa_ptr()), "slice of address and input slice don't match");
1562 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1563 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1564 debug_only(adr_type = C->get_adr_type(adr_idx));
1565 Node* mem = memory(adr_idx);
1566 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1567 ld = _gvn.transform(ld);
1568 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1569 // Improve graph before escape analysis and boxing elimination.
1570 record_for_igvn(ld);
1571 if (ld->is_DecodeN()) {
1572 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1573 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1574 // a Phi). Recording such cases is still perfectly sound, but may be
1575 // unnecessary and result in some minor IGVN overhead.
1576 record_for_igvn(ld->in(1));
1577 }
1578 }
1579 return ld;
1580 }
1581
1582 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1583 int adr_idx,
1584 MemNode::MemOrd mo,
1585 bool require_atomic_access,
1586 bool unaligned,
1587 bool mismatched,
1588 bool unsafe,
1589 int barrier_data) {
1590 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1591 assert(adr_idx == C->get_alias_index(_gvn.type(adr)->isa_ptr()), "slice of address and input slice don't match");
1592 const TypePtr* adr_type = nullptr;
1593 debug_only(adr_type = C->get_adr_type(adr_idx));
1594 Node *mem = memory(adr_idx);
1595 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1596 if (unaligned) {
1597 st->as_Store()->set_unaligned_access();
1598 }
1599 if (mismatched) {
1600 st->as_Store()->set_mismatched_access();
1601 }
1602 if (unsafe) {
1603 st->as_Store()->set_unsafe_access();
1604 }
1605 st->as_Store()->set_barrier_data(barrier_data);
1606 st = _gvn.transform(st);
1607 set_memory(st, adr_idx);
1608 // Back-to-back stores can only remove intermediate store with DU info
1609 // so push on worklist for optimizer.
1610 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1611 record_for_igvn(st);
1612
1613 return st;
1614 }
1615
1616 Node* GraphKit::access_store_at(Node* obj,
1617 Node* adr,
1618 const TypePtr* adr_type,
1619 Node* val,
1620 const Type* val_type,
1621 BasicType bt,
1622 DecoratorSet decorators) {
1623 // Transformation of a value which could be null pointer (CastPP #null)
1624 // could be delayed during Parse (for example, in adjust_map_after_if()).
1625 // Execute transformation here to avoid barrier generation in such case.
1626 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1627 val = _gvn.makecon(TypePtr::NULL_PTR);
1628 }
1629
1630 if (stopped()) {
1631 return top(); // Dead path ?
1632 }
1633
1634 assert(val != nullptr, "not dead path");
1635
1636 C2AccessValuePtr addr(adr, adr_type);
1637 C2AccessValue value(val, val_type);
1638 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1639 if (access.is_raw()) {
1640 return _barrier_set->BarrierSetC2::store_at(access, value);
1641 } else {
1642 return _barrier_set->store_at(access, value);
1643 }
1644 }
1645
1646 Node* GraphKit::access_load_at(Node* obj, // containing obj
1647 Node* adr, // actual address to store val at
1648 const TypePtr* adr_type,
1649 const Type* val_type,
1650 BasicType bt,
1651 DecoratorSet decorators) {
1652 if (stopped()) {
1653 return top(); // Dead path ?
1654 }
1655
1656 C2AccessValuePtr addr(adr, adr_type);
1657 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1658 if (access.is_raw()) {
1659 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1660 } else {
1661 return _barrier_set->load_at(access, val_type);
1662 }
1663 }
1664
1665 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1666 const Type* val_type,
1667 BasicType bt,
1668 DecoratorSet decorators) {
1669 if (stopped()) {
1670 return top(); // Dead path ?
1671 }
1672
1673 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1674 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1675 if (access.is_raw()) {
1676 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1677 } else {
1742 Node* new_val,
1743 const Type* value_type,
1744 BasicType bt,
1745 DecoratorSet decorators) {
1746 C2AccessValuePtr addr(adr, adr_type);
1747 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1748 if (access.is_raw()) {
1749 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1750 } else {
1751 return _barrier_set->atomic_add_at(access, new_val, value_type);
1752 }
1753 }
1754
1755 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1756 return _barrier_set->clone(this, src, dst, size, is_array);
1757 }
1758
1759 //-------------------------array_element_address-------------------------
1760 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1761 const TypeInt* sizetype, Node* ctrl) {
1762 uint shift = exact_log2(type2aelembytes(elembt));
1763 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1764
1765 // short-circuit a common case (saves lots of confusing waste motion)
1766 jint idx_con = find_int_con(idx, -1);
1767 if (idx_con >= 0) {
1768 intptr_t offset = header + ((intptr_t)idx_con << shift);
1769 return basic_plus_adr(ary, offset);
1770 }
1771
1772 // must be correct type for alignment purposes
1773 Node* base = basic_plus_adr(ary, header);
1774 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1775 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1776 return basic_plus_adr(ary, base, scale);
1777 }
1778
1779 //-------------------------load_array_element-------------------------
1780 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1781 const Type* elemtype = arytype->elem();
1782 BasicType elembt = elemtype->array_element_basic_type();
1783 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1784 if (elembt == T_NARROWOOP) {
1785 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1786 }
1787 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1788 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1789 return ld;
1790 }
1791
1792 //-------------------------set_arguments_for_java_call-------------------------
1793 // Arguments (pre-popped from the stack) are taken from the JVMS.
1794 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1795 // Add the call arguments:
1796 uint nargs = call->method()->arg_size();
1797 for (uint i = 0; i < nargs; i++) {
1798 Node* arg = argument(i);
1799 call->init_req(i + TypeFunc::Parms, arg);
1800 }
1801 }
1802
1803 //---------------------------set_edges_for_java_call---------------------------
1804 // Connect a newly created call into the current JVMS.
1805 // A return value node (if any) is returned from set_edges_for_java_call.
1806 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1807
1808 // Add the predefined inputs:
1809 call->init_req( TypeFunc::Control, control() );
1810 call->init_req( TypeFunc::I_O , i_o() );
1811 call->init_req( TypeFunc::Memory , reset_memory() );
1812 call->init_req( TypeFunc::FramePtr, frameptr() );
1813 call->init_req( TypeFunc::ReturnAdr, top() );
1814
1815 add_safepoint_edges(call, must_throw);
1816
1817 Node* xcall = _gvn.transform(call);
1818
1819 if (xcall == top()) {
1820 set_control(top());
1821 return;
1822 }
1823 assert(xcall == call, "call identity is stable");
1824
1825 // Re-use the current map to produce the result.
1826
1827 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1828 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1829 set_all_memory_call(xcall, separate_io_proj);
1830
1831 //return xcall; // no need, caller already has it
1832 }
1833
1834 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1835 if (stopped()) return top(); // maybe the call folded up?
1836
1837 // Capture the return value, if any.
1838 Node* ret;
1839 if (call->method() == nullptr ||
1840 call->method()->return_type()->basic_type() == T_VOID)
1841 ret = top();
1842 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1843
1844 // Note: Since any out-of-line call can produce an exception,
1845 // we always insert an I_O projection from the call into the result.
1846
1847 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1848
1849 if (separate_io_proj) {
1850 // The caller requested separate projections be used by the fall
1851 // through and exceptional paths, so replace the projections for
1852 // the fall through path.
1853 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1854 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1855 }
1856 return ret;
1857 }
1858
1859 //--------------------set_predefined_input_for_runtime_call--------------------
1860 // Reading and setting the memory state is way conservative here.
1861 // The real problem is that I am not doing real Type analysis on memory,
1862 // so I cannot distinguish card mark stores from other stores. Across a GC
1863 // point the Store Barrier and the card mark memory has to agree. I cannot
1864 // have a card mark store and its barrier split across the GC point from
1865 // either above or below. Here I get that to happen by reading ALL of memory.
1866 // A better answer would be to separate out card marks from other memory.
1867 // For now, return the input memory state, so that it can be reused
1868 // after the call, if this call has restricted memory effects.
1869 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1870 // Set fixed predefined input arguments
1871 Node* memory = reset_memory();
1872 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
1873 call->init_req( TypeFunc::Control, control() );
1874 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1875 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1926 if (use->is_MergeMem()) {
1927 wl.push(use);
1928 }
1929 }
1930 }
1931
1932 // Replace the call with the current state of the kit.
1933 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
1934 JVMState* ejvms = nullptr;
1935 if (has_exceptions()) {
1936 ejvms = transfer_exceptions_into_jvms();
1937 }
1938
1939 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1940 ReplacedNodes replaced_nodes_exception;
1941 Node* ex_ctl = top();
1942
1943 SafePointNode* final_state = stop();
1944
1945 // Find all the needed outputs of this call
1946 CallProjections callprojs;
1947 call->extract_projections(&callprojs, true, do_asserts);
1948
1949 Unique_Node_List wl;
1950 Node* init_mem = call->in(TypeFunc::Memory);
1951 Node* final_mem = final_state->in(TypeFunc::Memory);
1952 Node* final_ctl = final_state->in(TypeFunc::Control);
1953 Node* final_io = final_state->in(TypeFunc::I_O);
1954
1955 // Replace all the old call edges with the edges from the inlining result
1956 if (callprojs.fallthrough_catchproj != nullptr) {
1957 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1958 }
1959 if (callprojs.fallthrough_memproj != nullptr) {
1960 if (final_mem->is_MergeMem()) {
1961 // Parser's exits MergeMem was not transformed but may be optimized
1962 final_mem = _gvn.transform(final_mem);
1963 }
1964 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1965 add_mergemem_users_to_worklist(wl, final_mem);
1966 }
1967 if (callprojs.fallthrough_ioproj != nullptr) {
1968 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1969 }
1970
1971 // Replace the result with the new result if it exists and is used
1972 if (callprojs.resproj != nullptr && result != nullptr) {
1973 C->gvn_replace_by(callprojs.resproj, result);
1974 }
1975
1976 if (ejvms == nullptr) {
1977 // No exception edges to simply kill off those paths
1978 if (callprojs.catchall_catchproj != nullptr) {
1979 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1980 }
1981 if (callprojs.catchall_memproj != nullptr) {
1982 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1983 }
1984 if (callprojs.catchall_ioproj != nullptr) {
1985 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1986 }
1987 // Replace the old exception object with top
1988 if (callprojs.exobj != nullptr) {
1989 C->gvn_replace_by(callprojs.exobj, C->top());
1990 }
1991 } else {
1992 GraphKit ekit(ejvms);
1993
1994 // Load my combined exception state into the kit, with all phis transformed:
1995 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1996 replaced_nodes_exception = ex_map->replaced_nodes();
1997
1998 Node* ex_oop = ekit.use_exception_state(ex_map);
1999
2000 if (callprojs.catchall_catchproj != nullptr) {
2001 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2002 ex_ctl = ekit.control();
2003 }
2004 if (callprojs.catchall_memproj != nullptr) {
2005 Node* ex_mem = ekit.reset_memory();
2006 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2007 add_mergemem_users_to_worklist(wl, ex_mem);
2008 }
2009 if (callprojs.catchall_ioproj != nullptr) {
2010 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2011 }
2012
2013 // Replace the old exception object with the newly created one
2014 if (callprojs.exobj != nullptr) {
2015 C->gvn_replace_by(callprojs.exobj, ex_oop);
2016 }
2017 }
2018
2019 // Disconnect the call from the graph
2020 call->disconnect_inputs(C);
2021 C->gvn_replace_by(call, C->top());
2022
2023 // Clean up any MergeMems that feed other MergeMems since the
2024 // optimizer doesn't like that.
2025 while (wl.size() > 0) {
2026 _gvn.transform(wl.pop());
2027 }
2028
2029 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2030 replaced_nodes.apply(C, final_ctl);
2031 }
2032 if (!ex_ctl->is_top() && do_replaced_nodes) {
2033 replaced_nodes_exception.apply(C, ex_ctl);
2034 }
2035 }
2036
2037
2038 //------------------------------increment_counter------------------------------
2039 // for statistics: increment a VM counter by 1
2040
2041 void GraphKit::increment_counter(address counter_addr) {
2042 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2043 increment_counter(adr1);
2044 }
2045
2046 void GraphKit::increment_counter(Node* counter_addr) {
2047 int adr_type = Compile::AliasIdxRaw;
2048 Node* ctrl = control();
2049 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2210 *
2211 * @param n node that the type applies to
2212 * @param exact_kls type from profiling
2213 * @param maybe_null did profiling see null?
2214 *
2215 * @return node with improved type
2216 */
2217 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2218 const Type* current_type = _gvn.type(n);
2219 assert(UseTypeSpeculation, "type speculation must be on");
2220
2221 const TypePtr* speculative = current_type->speculative();
2222
2223 // Should the klass from the profile be recorded in the speculative type?
2224 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2225 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2226 const TypeOopPtr* xtype = tklass->as_instance_type();
2227 assert(xtype->klass_is_exact(), "Should be exact");
2228 // Any reason to believe n is not null (from this profiling or a previous one)?
2229 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2230 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2231 // record the new speculative type's depth
2232 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2233 speculative = speculative->with_inline_depth(jvms()->depth());
2234 } else if (current_type->would_improve_ptr(ptr_kind)) {
2235 // Profiling report that null was never seen so we can change the
2236 // speculative type to non null ptr.
2237 if (ptr_kind == ProfileAlwaysNull) {
2238 speculative = TypePtr::NULL_PTR;
2239 } else {
2240 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2241 const TypePtr* ptr = TypePtr::NOTNULL;
2242 if (speculative != nullptr) {
2243 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2244 } else {
2245 speculative = ptr;
2246 }
2247 }
2248 }
2249
2250 if (speculative != current_type->speculative()) {
2251 // Build a type with a speculative type (what we think we know
2252 // about the type but will need a guard when we use it)
2253 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2254 // We're changing the type, we need a new CheckCast node to carry
2255 // the new type. The new type depends on the control: what
2256 // profiling tells us is only valid from here as far as we can
2257 // tell.
2258 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2259 cast = _gvn.transform(cast);
2260 replace_in_map(n, cast);
2261 n = cast;
2262 }
2263
2264 return n;
2265 }
2266
2267 /**
2268 * Record profiling data from receiver profiling at an invoke with the
2269 * type system so that it can propagate it (speculation)
2270 *
2271 * @param n receiver node
2272 *
2273 * @return node with improved type
2274 */
2275 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2276 if (!UseTypeSpeculation) {
2277 return n;
2278 }
2279 ciKlass* exact_kls = profile_has_unique_klass();
2280 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2281 if ((java_bc() == Bytecodes::_checkcast ||
2282 java_bc() == Bytecodes::_instanceof ||
2283 java_bc() == Bytecodes::_aastore) &&
2284 method()->method_data()->is_mature()) {
2285 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2286 if (data != nullptr) {
2287 if (!data->as_BitData()->null_seen()) {
2288 ptr_kind = ProfileNeverNull;
2289 } else {
2290 assert(data->is_ReceiverTypeData(), "bad profile data type");
2291 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2292 uint i = 0;
2293 for (; i < call->row_limit(); i++) {
2294 ciKlass* receiver = call->receiver(i);
2295 if (receiver != nullptr) {
2296 break;
2297 }
2298 }
2299 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2300 }
2301 }
2302 }
2303 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2304 }
2305
2306 /**
2307 * Record profiling data from argument profiling at an invoke with the
2308 * type system so that it can propagate it (speculation)
2309 *
2310 * @param dest_method target method for the call
2311 * @param bc what invoke bytecode is this?
2312 */
2313 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2314 if (!UseTypeSpeculation) {
2315 return;
2316 }
2317 const TypeFunc* tf = TypeFunc::make(dest_method);
2318 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2319 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2320 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2321 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2322 if (is_reference_type(targ->basic_type())) {
2323 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2324 ciKlass* better_type = nullptr;
2325 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2326 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2327 }
2328 i++;
2329 }
2330 }
2331 }
2332
2333 /**
2334 * Record profiling data from parameter profiling at an invoke with
2335 * the type system so that it can propagate it (speculation)
2336 */
2337 void GraphKit::record_profiled_parameters_for_speculation() {
2338 if (!UseTypeSpeculation) {
2339 return;
2340 }
2341 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2355 * the type system so that it can propagate it (speculation)
2356 */
2357 void GraphKit::record_profiled_return_for_speculation() {
2358 if (!UseTypeSpeculation) {
2359 return;
2360 }
2361 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2362 ciKlass* better_type = nullptr;
2363 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2364 // If profiling reports a single type for the return value,
2365 // feed it to the type system so it can propagate it as a
2366 // speculative type
2367 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2368 }
2369 }
2370
2371 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2372 if (Matcher::strict_fp_requires_explicit_rounding) {
2373 // (Note: TypeFunc::make has a cache that makes this fast.)
2374 const TypeFunc* tf = TypeFunc::make(dest_method);
2375 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2376 for (int j = 0; j < nargs; j++) {
2377 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2378 if (targ->basic_type() == T_DOUBLE) {
2379 // If any parameters are doubles, they must be rounded before
2380 // the call, dprecision_rounding does gvn.transform
2381 Node *arg = argument(j);
2382 arg = dprecision_rounding(arg);
2383 set_argument(j, arg);
2384 }
2385 }
2386 }
2387 }
2388
2389 // rounding for strict float precision conformance
2390 Node* GraphKit::precision_rounding(Node* n) {
2391 if (Matcher::strict_fp_requires_explicit_rounding) {
2392 #ifdef IA32
2393 if (UseSSE == 0) {
2394 return _gvn.transform(new RoundFloatNode(nullptr, n));
2395 }
2396 #else
2397 Unimplemented();
2506 // The first null ends the list.
2507 Node* parm0, Node* parm1,
2508 Node* parm2, Node* parm3,
2509 Node* parm4, Node* parm5,
2510 Node* parm6, Node* parm7) {
2511 assert(call_addr != nullptr, "must not call null targets");
2512
2513 // Slow-path call
2514 bool is_leaf = !(flags & RC_NO_LEAF);
2515 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2516 if (call_name == nullptr) {
2517 assert(!is_leaf, "must supply name for leaf");
2518 call_name = OptoRuntime::stub_name(call_addr);
2519 }
2520 CallNode* call;
2521 if (!is_leaf) {
2522 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2523 } else if (flags & RC_NO_FP) {
2524 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2525 } else if (flags & RC_VECTOR){
2526 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2527 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2528 } else {
2529 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2530 }
2531
2532 // The following is similar to set_edges_for_java_call,
2533 // except that the memory effects of the call are restricted to AliasIdxRaw.
2534
2535 // Slow path call has no side-effects, uses few values
2536 bool wide_in = !(flags & RC_NARROW_MEM);
2537 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2538
2539 Node* prev_mem = nullptr;
2540 if (wide_in) {
2541 prev_mem = set_predefined_input_for_runtime_call(call);
2542 } else {
2543 assert(!wide_out, "narrow in => narrow out");
2544 Node* narrow_mem = memory(adr_type);
2545 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2546 }
2586
2587 if (has_io) {
2588 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2589 }
2590 return call;
2591
2592 }
2593
2594 // i2b
2595 Node* GraphKit::sign_extend_byte(Node* in) {
2596 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2597 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2598 }
2599
2600 // i2s
2601 Node* GraphKit::sign_extend_short(Node* in) {
2602 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2603 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2604 }
2605
2606 //------------------------------merge_memory-----------------------------------
2607 // Merge memory from one path into the current memory state.
2608 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2609 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2610 Node* old_slice = mms.force_memory();
2611 Node* new_slice = mms.memory2();
2612 if (old_slice != new_slice) {
2613 PhiNode* phi;
2614 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2615 if (mms.is_empty()) {
2616 // clone base memory Phi's inputs for this memory slice
2617 assert(old_slice == mms.base_memory(), "sanity");
2618 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2619 _gvn.set_type(phi, Type::MEMORY);
2620 for (uint i = 1; i < phi->req(); i++) {
2621 phi->init_req(i, old_slice->in(i));
2622 }
2623 } else {
2624 phi = old_slice->as_Phi(); // Phi was generated already
2625 }
2888
2889 // Now do a linear scan of the secondary super-klass array. Again, no real
2890 // performance impact (too rare) but it's gotta be done.
2891 // Since the code is rarely used, there is no penalty for moving it
2892 // out of line, and it can only improve I-cache density.
2893 // The decision to inline or out-of-line this final check is platform
2894 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2895 Node* psc = gvn.transform(
2896 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2897
2898 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2899 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2900 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2901
2902 // Return false path; set default control to true path.
2903 *ctrl = gvn.transform(r_ok_subtype);
2904 return gvn.transform(r_not_subtype);
2905 }
2906
2907 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2908 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2909 if (expand_subtype_check) {
2910 MergeMemNode* mem = merged_memory();
2911 Node* ctrl = control();
2912 Node* subklass = obj_or_subklass;
2913 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2914 subklass = load_object_klass(obj_or_subklass);
2915 }
2916
2917 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2918 set_control(ctrl);
2919 return n;
2920 }
2921
2922 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2923 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2924 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2925 set_control(_gvn.transform(new IfTrueNode(iff)));
2926 return _gvn.transform(new IfFalseNode(iff));
2927 }
2928
2929 // Profile-driven exact type check:
2930 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2931 float prob,
2932 Node* *casted_receiver) {
2933 assert(!klass->is_interface(), "no exact type check on interfaces");
2934
2935 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
2936 Node* recv_klass = load_object_klass(receiver);
2937 Node* want_klass = makecon(tklass);
2938 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
2939 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
2940 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2941 set_control( _gvn.transform(new IfTrueNode (iff)));
2942 Node* fail = _gvn.transform(new IfFalseNode(iff));
2943
2944 if (!stopped()) {
2945 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2946 const TypeOopPtr* recvx_type = tklass->as_instance_type();
2947 assert(recvx_type->klass_is_exact(), "");
2948
2949 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
2950 // Subsume downstream occurrences of receiver with a cast to
2951 // recv_xtype, since now we know what the type will be.
2952 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
2953 (*casted_receiver) = _gvn.transform(cast);
2954 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
2955 // (User must make the replace_in_map call.)
2956 }
2957 }
2958
2959 return fail;
2960 }
2961
2962 //------------------------------subtype_check_receiver-------------------------
2963 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2964 Node** casted_receiver) {
2965 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
2966 Node* want_klass = makecon(tklass);
2967
2968 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2969
2970 // Ignore interface type information until interface types are properly tracked.
2971 if (!stopped() && !klass->is_interface()) {
2972 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2973 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2974 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
2975 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2976 (*casted_receiver) = _gvn.transform(cast);
2977 }
2978 }
2979
2980 return slow_ctl;
2981 }
2982
2983 //------------------------------seems_never_null-------------------------------
2984 // Use null_seen information if it is available from the profile.
2985 // If we see an unexpected null at a type check we record it and force a
2986 // recompile; the offending check will be recompiled to handle nulls.
2987 // If we see several offending BCIs, then all checks in the
2988 // method will be recompiled.
2989 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2990 speculating = !_gvn.type(obj)->speculative_maybe_null();
2991 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2992 if (UncommonNullCast // Cutout for this technique
2993 && obj != null() // And not the -Xcomp stupid case?
2994 && !too_many_traps(reason)
2995 ) {
2996 if (speculating) {
3065
3066 //------------------------maybe_cast_profiled_receiver-------------------------
3067 // If the profile has seen exactly one type, narrow to exactly that type.
3068 // Subsequent type checks will always fold up.
3069 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3070 const TypeKlassPtr* require_klass,
3071 ciKlass* spec_klass,
3072 bool safe_for_replace) {
3073 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3074
3075 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3076
3077 // Make sure we haven't already deoptimized from this tactic.
3078 if (too_many_traps_or_recompiles(reason))
3079 return nullptr;
3080
3081 // (No, this isn't a call, but it's enough like a virtual call
3082 // to use the same ciMethod accessor to get the profile info...)
3083 // If we have a speculative type use it instead of profiling (which
3084 // may not help us)
3085 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3086 if (exact_kls != nullptr) {// no cast failures here
3087 if (require_klass == nullptr ||
3088 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3089 // If we narrow the type to match what the type profile sees or
3090 // the speculative type, we can then remove the rest of the
3091 // cast.
3092 // This is a win, even if the exact_kls is very specific,
3093 // because downstream operations, such as method calls,
3094 // will often benefit from the sharper type.
3095 Node* exact_obj = not_null_obj; // will get updated in place...
3096 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3097 &exact_obj);
3098 { PreserveJVMState pjvms(this);
3099 set_control(slow_ctl);
3100 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3101 }
3102 if (safe_for_replace) {
3103 replace_in_map(not_null_obj, exact_obj);
3104 }
3105 return exact_obj;
3195 // If not_null_obj is dead, only null-path is taken
3196 if (stopped()) { // Doing instance-of on a null?
3197 set_control(null_ctl);
3198 return intcon(0);
3199 }
3200 region->init_req(_null_path, null_ctl);
3201 phi ->init_req(_null_path, intcon(0)); // Set null path value
3202 if (null_ctl == top()) {
3203 // Do this eagerly, so that pattern matches like is_diamond_phi
3204 // will work even during parsing.
3205 assert(_null_path == PATH_LIMIT-1, "delete last");
3206 region->del_req(_null_path);
3207 phi ->del_req(_null_path);
3208 }
3209
3210 // Do we know the type check always succeed?
3211 bool known_statically = false;
3212 if (_gvn.type(superklass)->singleton()) {
3213 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3214 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3215 if (subk->is_loaded()) {
3216 int static_res = C->static_subtype_check(superk, subk);
3217 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3218 }
3219 }
3220
3221 if (!known_statically) {
3222 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3223 // We may not have profiling here or it may not help us. If we
3224 // have a speculative type use it to perform an exact cast.
3225 ciKlass* spec_obj_type = obj_type->speculative_type();
3226 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3227 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3228 if (stopped()) { // Profile disagrees with this path.
3229 set_control(null_ctl); // Null is the only remaining possibility.
3230 return intcon(0);
3231 }
3232 if (cast_obj != nullptr) {
3233 not_null_obj = cast_obj;
3234 }
3235 }
3251 record_for_igvn(region);
3252
3253 // If we know the type check always succeeds then we don't use the
3254 // profiling data at this bytecode. Don't lose it, feed it to the
3255 // type system as a speculative type.
3256 if (safe_for_replace) {
3257 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3258 replace_in_map(obj, casted_obj);
3259 }
3260
3261 return _gvn.transform(phi);
3262 }
3263
3264 //-------------------------------gen_checkcast---------------------------------
3265 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3266 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3267 // uncommon-trap paths work. Adjust stack after this call.
3268 // If failure_control is supplied and not null, it is filled in with
3269 // the control edge for the cast failure. Otherwise, an appropriate
3270 // uncommon trap or exception is thrown.
3271 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3272 Node* *failure_control) {
3273 kill_dead_locals(); // Benefit all the uncommon traps
3274 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3275 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3276 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3277
3278 // Fast cutout: Check the case that the cast is vacuously true.
3279 // This detects the common cases where the test will short-circuit
3280 // away completely. We do this before we perform the null check,
3281 // because if the test is going to turn into zero code, we don't
3282 // want a residual null check left around. (Causes a slowdown,
3283 // for example, in some objArray manipulations, such as a[i]=a[j].)
3284 if (improved_klass_ptr_type->singleton()) {
3285 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3286 if (objtp != nullptr) {
3287 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3288 case Compile::SSC_always_true:
3289 // If we know the type check always succeed then we don't use
3290 // the profiling data at this bytecode. Don't lose it, feed it
3291 // to the type system as a speculative type.
3292 return record_profiled_receiver_for_speculation(obj);
3293 case Compile::SSC_always_false:
3294 // It needs a null check because a null will *pass* the cast check.
3295 // A non-null value will always produce an exception.
3296 if (!objtp->maybe_null()) {
3297 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3298 Deoptimization::DeoptReason reason = is_aastore ?
3299 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3300 builtin_throw(reason);
3301 return top();
3302 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3303 return null_assert(obj);
3304 }
3305 break; // Fall through to full check
3306 default:
3307 break;
3308 }
3309 }
3310 }
3311
3312 ciProfileData* data = nullptr;
3313 bool safe_for_replace = false;
3314 if (failure_control == nullptr) { // use MDO in regular case only
3315 assert(java_bc() == Bytecodes::_aastore ||
3316 java_bc() == Bytecodes::_checkcast,
3317 "interpreter profiles type checks only for these BCs");
3318 data = method()->method_data()->bci_to_data(bci());
3319 safe_for_replace = true;
3320 }
3321
3322 // Make the merge point
3323 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3324 RegionNode* region = new RegionNode(PATH_LIMIT);
3325 Node* phi = new PhiNode(region, toop);
3326 C->set_has_split_ifs(true); // Has chance for split-if optimization
3327
3328 // Use null-cast information if it is available
3329 bool speculative_not_null = false;
3330 bool never_see_null = ((failure_control == nullptr) // regular case only
3331 && seems_never_null(obj, data, speculative_not_null));
3332
3333 // Null check; get casted pointer; set region slot 3
3334 Node* null_ctl = top();
3335 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3336
3337 // If not_null_obj is dead, only null-path is taken
3338 if (stopped()) { // Doing instance-of on a null?
3339 set_control(null_ctl);
3340 return null();
3341 }
3342 region->init_req(_null_path, null_ctl);
3343 phi ->init_req(_null_path, null()); // Set null path value
3344 if (null_ctl == top()) {
3345 // Do this eagerly, so that pattern matches like is_diamond_phi
3346 // will work even during parsing.
3347 assert(_null_path == PATH_LIMIT-1, "delete last");
3348 region->del_req(_null_path);
3349 phi ->del_req(_null_path);
3350 }
3351
3352 Node* cast_obj = nullptr;
3353 if (improved_klass_ptr_type->klass_is_exact()) {
3354 // The following optimization tries to statically cast the speculative type of the object
3355 // (for example obtained during profiling) to the type of the superklass and then do a
3356 // dynamic check that the type of the object is what we expect. To work correctly
3357 // for checkcast and aastore the type of superklass should be exact.
3358 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3359 // We may not have profiling here or it may not help us. If we have
3360 // a speculative type use it to perform an exact cast.
3361 ciKlass* spec_obj_type = obj_type->speculative_type();
3362 if (spec_obj_type != nullptr || data != nullptr) {
3363 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3364 if (cast_obj != nullptr) {
3365 if (failure_control != nullptr) // failure is now impossible
3366 (*failure_control) = top();
3367 // adjust the type of the phi to the exact klass:
3368 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3369 }
3370 }
3371 }
3372
3373 if (cast_obj == nullptr) {
3374 // Generate the subtype check
3375 Node* improved_superklass = superklass;
3376 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3377 improved_superklass = makecon(improved_klass_ptr_type);
3378 }
3379 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3380
3381 // Plug in success path into the merge
3382 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3383 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3384 if (failure_control == nullptr) {
3385 if (not_subtype_ctrl != top()) { // If failure is possible
3386 PreserveJVMState pjvms(this);
3387 set_control(not_subtype_ctrl);
3388 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3389 Deoptimization::DeoptReason reason = is_aastore ?
3390 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3391 builtin_throw(reason);
3392 }
3393 } else {
3394 (*failure_control) = not_subtype_ctrl;
3395 }
3396 }
3397
3398 region->init_req(_obj_path, control());
3399 phi ->init_req(_obj_path, cast_obj);
3400
3401 // A merge of null or Casted-NotNull obj
3402 Node* res = _gvn.transform(phi);
3403
3404 // Note I do NOT always 'replace_in_map(obj,result)' here.
3405 // if( tk->klass()->can_be_primary_super() )
3406 // This means that if I successfully store an Object into an array-of-String
3407 // I 'forget' that the Object is really now known to be a String. I have to
3408 // do this because we don't have true union types for interfaces - if I store
3409 // a Baz into an array-of-Interface and then tell the optimizer it's an
3410 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3411 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3412 // replace_in_map( obj, res );
3413
3414 // Return final merged results
3415 set_control( _gvn.transform(region) );
3416 record_for_igvn(region);
3417
3418 return record_profiled_receiver_for_speculation(res);
3419 }
3420
3421 //------------------------------next_monitor-----------------------------------
3422 // What number should be given to the next monitor?
3423 int GraphKit::next_monitor() {
3424 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3425 int next = current + C->sync_stack_slots();
3426 // Keep the toplevel high water mark current:
3427 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3428 return current;
3429 }
3430
3431 //------------------------------insert_mem_bar---------------------------------
3432 // Memory barrier to avoid floating things around
3433 // The membar serves as a pinch point between both control and all memory slices.
3434 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3435 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3436 mb->init_req(TypeFunc::Control, control());
3437 mb->init_req(TypeFunc::Memory, reset_memory());
3438 Node* membar = _gvn.transform(mb);
3466 }
3467 Node* membar = _gvn.transform(mb);
3468 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3469 if (alias_idx == Compile::AliasIdxBot) {
3470 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3471 } else {
3472 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3473 }
3474 return membar;
3475 }
3476
3477 //------------------------------shared_lock------------------------------------
3478 // Emit locking code.
3479 FastLockNode* GraphKit::shared_lock(Node* obj) {
3480 // bci is either a monitorenter bc or InvocationEntryBci
3481 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3482 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3483
3484 if( !GenerateSynchronizationCode )
3485 return nullptr; // Not locking things?
3486 if (stopped()) // Dead monitor?
3487 return nullptr;
3488
3489 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3490
3491 // Box the stack location
3492 Node* box = new BoxLockNode(next_monitor());
3493 // Check for bailout after new BoxLockNode
3494 if (failing()) { return nullptr; }
3495 box = _gvn.transform(box);
3496 Node* mem = reset_memory();
3497
3498 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3499
3500 // Add monitor to debug info for the slow path. If we block inside the
3501 // slow path and de-opt, we need the monitor hanging around
3502 map()->push_monitor( flock );
3503
3504 const TypeFunc *tf = LockNode::lock_type();
3505 LockNode *lock = new LockNode(C, tf);
3534 }
3535 #endif
3536
3537 return flock;
3538 }
3539
3540
3541 //------------------------------shared_unlock----------------------------------
3542 // Emit unlocking code.
3543 void GraphKit::shared_unlock(Node* box, Node* obj) {
3544 // bci is either a monitorenter bc or InvocationEntryBci
3545 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3546 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3547
3548 if( !GenerateSynchronizationCode )
3549 return;
3550 if (stopped()) { // Dead monitor?
3551 map()->pop_monitor(); // Kill monitor from debug info
3552 return;
3553 }
3554
3555 // Memory barrier to avoid floating things down past the locked region
3556 insert_mem_bar(Op_MemBarReleaseLock);
3557
3558 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3559 UnlockNode *unlock = new UnlockNode(C, tf);
3560 #ifdef ASSERT
3561 unlock->set_dbg_jvms(sync_jvms());
3562 #endif
3563 uint raw_idx = Compile::AliasIdxRaw;
3564 unlock->init_req( TypeFunc::Control, control() );
3565 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3566 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3567 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3568 unlock->init_req( TypeFunc::ReturnAdr, top() );
3569
3570 unlock->init_req(TypeFunc::Parms + 0, obj);
3571 unlock->init_req(TypeFunc::Parms + 1, box);
3572 unlock = _gvn.transform(unlock)->as_Unlock();
3573
3574 Node* mem = reset_memory();
3575
3576 // unlock has no side-effects, sets few values
3577 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3578
3579 // Kill monitor from debug info
3580 map()->pop_monitor( );
3581 }
3582
3583 //-------------------------------get_layout_helper-----------------------------
3584 // If the given klass is a constant or known to be an array,
3585 // fetch the constant layout helper value into constant_value
3586 // and return null. Otherwise, load the non-constant
3587 // layout helper value, and return the node which represents it.
3588 // This two-faced routine is useful because allocation sites
3589 // almost always feature constant types.
3590 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3591 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3592 if (!StressReflectiveCode && klass_t != nullptr) {
3593 bool xklass = klass_t->klass_is_exact();
3594 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3595 jint lhelper;
3596 if (klass_t->isa_aryklassptr()) {
3597 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3598 if (is_reference_type(elem, true)) {
3599 elem = T_OBJECT;
3600 }
3601 lhelper = Klass::array_layout_helper(elem);
3602 } else {
3603 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3604 }
3605 if (lhelper != Klass::_lh_neutral_value) {
3606 constant_value = lhelper;
3607 return (Node*) nullptr;
3608 }
3609 }
3610 }
3611 constant_value = Klass::_lh_neutral_value; // put in a known value
3612 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3613 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3614 }
3615
3616 // We just put in an allocate/initialize with a big raw-memory effect.
3617 // Hook selected additional alias categories on the initialization.
3618 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3619 MergeMemNode* init_in_merge,
3620 Node* init_out_raw) {
3621 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3622 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3623
3624 Node* prevmem = kit.memory(alias_idx);
3625 init_in_merge->set_memory_at(alias_idx, prevmem);
3626 kit.set_memory(init_out_raw, alias_idx);
3627 }
3628
3629 //---------------------------set_output_for_allocation-------------------------
3630 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3631 const TypeOopPtr* oop_type,
3632 bool deoptimize_on_exception) {
3633 int rawidx = Compile::AliasIdxRaw;
3634 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3635 add_safepoint_edges(alloc);
3636 Node* allocx = _gvn.transform(alloc);
3637 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3638 // create memory projection for i_o
3639 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3640 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3641
3642 // create a memory projection as for the normal control path
3643 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3644 set_memory(malloc, rawidx);
3645
3646 // a normal slow-call doesn't change i_o, but an allocation does
3647 // we create a separate i_o projection for the normal control path
3648 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3649 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3650
3651 // put in an initialization barrier
3652 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3653 rawoop)->as_Initialize();
3654 assert(alloc->initialization() == init, "2-way macro link must work");
3655 assert(init ->allocation() == alloc, "2-way macro link must work");
3656 {
3657 // Extract memory strands which may participate in the new object's
3658 // initialization, and source them from the new InitializeNode.
3659 // This will allow us to observe initializations when they occur,
3660 // and link them properly (as a group) to the InitializeNode.
3661 assert(init->in(InitializeNode::Memory) == malloc, "");
3662 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3663 init->set_req(InitializeNode::Memory, minit_in);
3664 record_for_igvn(minit_in); // fold it up later, if possible
3665 Node* minit_out = memory(rawidx);
3666 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3667 // Add an edge in the MergeMem for the header fields so an access
3668 // to one of those has correct memory state
3669 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3670 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3671 if (oop_type->isa_aryptr()) {
3672 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3673 int elemidx = C->get_alias_index(telemref);
3674 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3675 } else if (oop_type->isa_instptr()) {
3676 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3677 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3678 ciField* field = ik->nonstatic_field_at(i);
3679 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3680 continue; // do not bother to track really large numbers of fields
3681 // Find (or create) the alias category for this field:
3682 int fieldidx = C->alias_type(field)->index();
3683 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3684 }
3685 }
3686 }
3687
3688 // Cast raw oop to the real thing...
3689 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3690 javaoop = _gvn.transform(javaoop);
3691 C->set_recent_alloc(control(), javaoop);
3692 assert(just_allocated_object(control()) == javaoop, "just allocated");
3693
3694 #ifdef ASSERT
3695 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3706 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3707 }
3708 }
3709 #endif //ASSERT
3710
3711 return javaoop;
3712 }
3713
3714 //---------------------------new_instance--------------------------------------
3715 // This routine takes a klass_node which may be constant (for a static type)
3716 // or may be non-constant (for reflective code). It will work equally well
3717 // for either, and the graph will fold nicely if the optimizer later reduces
3718 // the type to a constant.
3719 // The optional arguments are for specialized use by intrinsics:
3720 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3721 // - If 'return_size_val', report the total object size to the caller.
3722 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3723 Node* GraphKit::new_instance(Node* klass_node,
3724 Node* extra_slow_test,
3725 Node* *return_size_val,
3726 bool deoptimize_on_exception) {
3727 // Compute size in doublewords
3728 // The size is always an integral number of doublewords, represented
3729 // as a positive bytewise size stored in the klass's layout_helper.
3730 // The layout_helper also encodes (in a low bit) the need for a slow path.
3731 jint layout_con = Klass::_lh_neutral_value;
3732 Node* layout_val = get_layout_helper(klass_node, layout_con);
3733 int layout_is_con = (layout_val == nullptr);
3734
3735 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3736 // Generate the initial go-slow test. It's either ALWAYS (return a
3737 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3738 // case) a computed value derived from the layout_helper.
3739 Node* initial_slow_test = nullptr;
3740 if (layout_is_con) {
3741 assert(!StressReflectiveCode, "stress mode does not use these paths");
3742 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3743 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3744 } else { // reflective case
3745 // This reflective path is used by Unsafe.allocateInstance.
3746 // (It may be stress-tested by specifying StressReflectiveCode.)
3747 // Basically, we want to get into the VM is there's an illegal argument.
3748 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3749 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3750 if (extra_slow_test != intcon(0)) {
3751 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3752 }
3753 // (Macro-expander will further convert this to a Bool, if necessary.)
3764
3765 // Clear the low bits to extract layout_helper_size_in_bytes:
3766 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3767 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3768 size = _gvn.transform( new AndXNode(size, mask) );
3769 }
3770 if (return_size_val != nullptr) {
3771 (*return_size_val) = size;
3772 }
3773
3774 // This is a precise notnull oop of the klass.
3775 // (Actually, it need not be precise if this is a reflective allocation.)
3776 // It's what we cast the result to.
3777 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3778 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3779 const TypeOopPtr* oop_type = tklass->as_instance_type();
3780
3781 // Now generate allocation code
3782
3783 // The entire memory state is needed for slow path of the allocation
3784 // since GC and deoptimization can happened.
3785 Node *mem = reset_memory();
3786 set_all_memory(mem); // Create new memory state
3787
3788 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3789 control(), mem, i_o(),
3790 size, klass_node,
3791 initial_slow_test);
3792
3793 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3794 }
3795
3796 //-------------------------------new_array-------------------------------------
3797 // helper for both newarray and anewarray
3798 // The 'length' parameter is (obviously) the length of the array.
3799 // The optional arguments are for specialized use by intrinsics:
3800 // - If 'return_size_val', report the non-padded array size (sum of header size
3801 // and array body) to the caller.
3802 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3803 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3804 Node* length, // number of array elements
3805 int nargs, // number of arguments to push back for uncommon trap
3806 Node* *return_size_val,
3807 bool deoptimize_on_exception) {
3808 jint layout_con = Klass::_lh_neutral_value;
3809 Node* layout_val = get_layout_helper(klass_node, layout_con);
3810 int layout_is_con = (layout_val == nullptr);
3811
3812 if (!layout_is_con && !StressReflectiveCode &&
3813 !too_many_traps(Deoptimization::Reason_class_check)) {
3814 // This is a reflective array creation site.
3815 // Optimistically assume that it is a subtype of Object[],
3816 // so that we can fold up all the address arithmetic.
3817 layout_con = Klass::array_layout_helper(T_OBJECT);
3818 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3819 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3820 { BuildCutout unless(this, bol_lh, PROB_MAX);
3821 inc_sp(nargs);
3822 uncommon_trap(Deoptimization::Reason_class_check,
3823 Deoptimization::Action_maybe_recompile);
3824 }
3825 layout_val = nullptr;
3826 layout_is_con = true;
3827 }
3828
3829 // Generate the initial go-slow test. Make sure we do not overflow
3830 // if length is huge (near 2Gig) or negative! We do not need
3831 // exact double-words here, just a close approximation of needed
3832 // double-words. We can't add any offset or rounding bits, lest we
3833 // take a size -1 of bytes and make it positive. Use an unsigned
3834 // compare, so negative sizes look hugely positive.
3835 int fast_size_limit = FastAllocateSizeLimit;
3836 if (layout_is_con) {
3837 assert(!StressReflectiveCode, "stress mode does not use these paths");
3838 // Increase the size limit if we have exact knowledge of array type.
3839 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3840 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3841 }
3842
3843 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3844 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3845
3846 // --- Size Computation ---
3847 // array_size = round_to_heap(array_header + (length << elem_shift));
3848 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3849 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3850 // The rounding mask is strength-reduced, if possible.
3851 int round_mask = MinObjAlignmentInBytes - 1;
3852 Node* header_size = nullptr;
3853 // (T_BYTE has the weakest alignment and size restrictions...)
3854 if (layout_is_con) {
3855 int hsize = Klass::layout_helper_header_size(layout_con);
3856 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3857 if ((round_mask & ~right_n_bits(eshift)) == 0)
3858 round_mask = 0; // strength-reduce it if it goes away completely
3859 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3860 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3861 assert(header_size_min <= hsize, "generic minimum is smallest");
3862 header_size = intcon(hsize);
3863 } else {
3864 Node* hss = intcon(Klass::_lh_header_size_shift);
3865 Node* hsm = intcon(Klass::_lh_header_size_mask);
3866 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3867 header_size = _gvn.transform(new AndINode(header_size, hsm));
3868 }
3869
3870 Node* elem_shift = nullptr;
3871 if (layout_is_con) {
3872 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3873 if (eshift != 0)
3874 elem_shift = intcon(eshift);
3875 } else {
3876 // There is no need to mask or shift this value.
3877 // The semantics of LShiftINode include an implicit mask to 0x1F.
3878 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3879 elem_shift = layout_val;
3926 }
3927 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
3928
3929 if (return_size_val != nullptr) {
3930 // This is the size
3931 (*return_size_val) = non_rounded_size;
3932 }
3933
3934 Node* size = non_rounded_size;
3935 if (round_mask != 0) {
3936 Node* mask1 = MakeConX(round_mask);
3937 size = _gvn.transform(new AddXNode(size, mask1));
3938 Node* mask2 = MakeConX(~round_mask);
3939 size = _gvn.transform(new AndXNode(size, mask2));
3940 }
3941 // else if round_mask == 0, the size computation is self-rounding
3942
3943 // Now generate allocation code
3944
3945 // The entire memory state is needed for slow path of the allocation
3946 // since GC and deoptimization can happened.
3947 Node *mem = reset_memory();
3948 set_all_memory(mem); // Create new memory state
3949
3950 if (initial_slow_test->is_Bool()) {
3951 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3952 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3953 }
3954
3955 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3956 Node* valid_length_test = _gvn.intcon(1);
3957 if (ary_type->isa_aryptr()) {
3958 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
3959 jint max = TypeAryPtr::max_array_length(bt);
3960 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
3961 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
3962 }
3963
3964 // Create the AllocateArrayNode and its result projections
3965 AllocateArrayNode* alloc
3966 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3967 control(), mem, i_o(),
3968 size, klass_node,
3969 initial_slow_test,
3970 length, valid_length_test);
3971
3972 // Cast to correct type. Note that the klass_node may be constant or not,
3973 // and in the latter case the actual array type will be inexact also.
3974 // (This happens via a non-constant argument to inline_native_newArray.)
3975 // In any case, the value of klass_node provides the desired array type.
3976 const TypeInt* length_type = _gvn.find_int_type(length);
3977 if (ary_type->isa_aryptr() && length_type != nullptr) {
3978 // Try to get a better type than POS for the size
3979 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3980 }
3981
3982 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3983
3984 array_ideal_length(alloc, ary_type, true);
3985 return javaoop;
3986 }
3987
3988 // The following "Ideal_foo" functions are placed here because they recognize
3989 // the graph shapes created by the functions immediately above.
3990
3991 //---------------------------Ideal_allocation----------------------------------
4098 set_all_memory(ideal.merged_memory());
4099 set_i_o(ideal.i_o());
4100 set_control(ideal.ctrl());
4101 }
4102
4103 void GraphKit::final_sync(IdealKit& ideal) {
4104 // Final sync IdealKit and graphKit.
4105 sync_kit(ideal);
4106 }
4107
4108 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4109 Node* len = load_array_length(load_String_value(str, set_ctrl));
4110 Node* coder = load_String_coder(str, set_ctrl);
4111 // Divide length by 2 if coder is UTF16
4112 return _gvn.transform(new RShiftINode(len, coder));
4113 }
4114
4115 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4116 int value_offset = java_lang_String::value_offset();
4117 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4118 false, nullptr, 0);
4119 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4120 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4121 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4122 ciTypeArrayKlass::make(T_BYTE), true, 0);
4123 Node* p = basic_plus_adr(str, str, value_offset);
4124 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4125 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4126 return load;
4127 }
4128
4129 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4130 if (!CompactStrings) {
4131 return intcon(java_lang_String::CODER_UTF16);
4132 }
4133 int coder_offset = java_lang_String::coder_offset();
4134 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4135 false, nullptr, 0);
4136 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4137
4138 Node* p = basic_plus_adr(str, str, coder_offset);
4139 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4140 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4141 return load;
4142 }
4143
4144 void GraphKit::store_String_value(Node* str, Node* value) {
4145 int value_offset = java_lang_String::value_offset();
4146 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4147 false, nullptr, 0);
4148 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4149
4150 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4151 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4152 }
4153
4154 void GraphKit::store_String_coder(Node* str, Node* value) {
4155 int coder_offset = java_lang_String::coder_offset();
4156 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4157 false, nullptr, 0);
4158 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4159
4160 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4161 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4162 }
4163
4164 // Capture src and dst memory state with a MergeMemNode
4165 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4166 if (src_type == dst_type) {
4167 // Types are equal, we don't need a MergeMemNode
4168 return memory(src_type);
4169 }
4170 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4171 record_for_igvn(merge); // fold it up later, if possible
4172 int src_idx = C->get_alias_index(src_type);
4173 int dst_idx = C->get_alias_index(dst_type);
4174 merge->set_memory_at(src_idx, memory(src_idx));
4175 merge->set_memory_at(dst_idx, memory(dst_idx));
4176 return merge;
4177 }
4250 i_char->init_req(2, AddI(i_char, intcon(2)));
4251
4252 set_control(IfFalse(iff));
4253 set_memory(st, TypeAryPtr::BYTES);
4254 }
4255
4256 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4257 if (!field->is_constant()) {
4258 return nullptr; // Field not marked as constant.
4259 }
4260 ciInstance* holder = nullptr;
4261 if (!field->is_static()) {
4262 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4263 if (const_oop != nullptr && const_oop->is_instance()) {
4264 holder = const_oop->as_instance();
4265 }
4266 }
4267 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4268 /*is_unsigned_load=*/false);
4269 if (con_type != nullptr) {
4270 return makecon(con_type);
4271 }
4272 return nullptr;
4273 }
4274
4275 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4276 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4277 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4278 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4279 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4280 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4281 return casted_obj;
4282 }
4283 return obj;
4284 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciUtilities.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "ci/ciObjArray.hpp"
31 #include "asm/register.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/inlinetypenode.hpp"
43 #include "opto/intrinsicnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "runtime/deoptimization.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "utilities/bitMap.inline.hpp"
55 #include "utilities/powerOfTwo.hpp"
56 #include "utilities/growableArray.hpp"
57
58 //----------------------------GraphKit-----------------------------------------
59 // Main utility constructor.
60 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
61 : Phase(Phase::Parser),
62 _env(C->env()),
63 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
64 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
65 {
66 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
67 _exceptions = jvms->map()->next_exception();
68 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
69 set_jvms(jvms);
70 #ifdef ASSERT
71 if (_gvn.is_IterGVN() != nullptr) {
72 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
73 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
74 _worklist_size = _gvn.C->igvn_worklist()->size();
75 }
76 #endif
77 }
78
79 // Private constructor for parser.
80 GraphKit::GraphKit()
81 : Phase(Phase::Parser),
82 _env(C->env()),
83 _gvn(*C->initial_gvn()),
84 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
85 {
86 _exceptions = nullptr;
87 set_map(nullptr);
88 debug_only(_sp = -99);
89 debug_only(set_bci(-99));
90 }
91
92
93
94 //---------------------------clean_stack---------------------------------------
95 // Clear away rubbish from the stack area of the JVM state.
96 // This destroys any arguments that may be waiting on the stack.
854 if (PrintMiscellaneous && (Verbose || WizardMode)) {
855 tty->print_cr("Zombie local %d: ", local);
856 jvms->dump();
857 }
858 return false;
859 }
860 }
861 }
862 return true;
863 }
864
865 #endif //ASSERT
866
867 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
868 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
869 ciMethod* cur_method = jvms->method();
870 int cur_bci = jvms->bci();
871 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
872 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
873 return Interpreter::bytecode_should_reexecute(code) ||
874 (is_anewarray && (code == Bytecodes::_multianewarray));
875 // Reexecute _multianewarray bytecode which was replaced with
876 // sequence of [a]newarray. See Parse::do_multianewarray().
877 //
878 // Note: interpreter should not have it set since this optimization
879 // is limited by dimensions and guarded by flag so in some cases
880 // multianewarray() runtime calls will be generated and
881 // the bytecode should not be reexecutes (stack will not be reset).
882 } else {
883 return false;
884 }
885 }
886
887 // Helper function for adding JVMState and debug information to node
888 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
889 // Add the safepoint edges to the call (or other safepoint).
890
891 // Make sure dead locals are set to top. This
892 // should help register allocation time and cut down on the size
893 // of the deoptimization information.
894 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
945 }
946
947 // Presize the call:
948 DEBUG_ONLY(uint non_debug_edges = call->req());
949 call->add_req_batch(top(), youngest_jvms->debug_depth());
950 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
951
952 // Set up edges so that the call looks like this:
953 // Call [state:] ctl io mem fptr retadr
954 // [parms:] parm0 ... parmN
955 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
956 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
957 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
958 // Note that caller debug info precedes callee debug info.
959
960 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
961 uint debug_ptr = call->req();
962
963 // Loop over the map input edges associated with jvms, add them
964 // to the call node, & reset all offsets to match call node array.
965
966 JVMState* callee_jvms = nullptr;
967 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
968 uint debug_end = debug_ptr;
969 uint debug_start = debug_ptr - in_jvms->debug_size();
970 debug_ptr = debug_start; // back up the ptr
971
972 uint p = debug_start; // walks forward in [debug_start, debug_end)
973 uint j, k, l;
974 SafePointNode* in_map = in_jvms->map();
975 out_jvms->set_map(call);
976
977 if (can_prune_locals) {
978 assert(in_jvms->method() == out_jvms->method(), "sanity");
979 // If the current throw can reach an exception handler in this JVMS,
980 // then we must keep everything live that can reach that handler.
981 // As a quick and dirty approximation, we look for any handlers at all.
982 if (in_jvms->method()->has_exception_handlers()) {
983 can_prune_locals = false;
984 }
985 }
986
987 // Add the Locals
988 k = in_jvms->locoff();
989 l = in_jvms->loc_size();
990 out_jvms->set_locoff(p);
991 if (!can_prune_locals) {
992 for (j = 0; j < l; j++) {
993 Node* val = in_map->in(k + j);
994 // Check if there's a larval that has been written in the callee state (constructor) and update it in the caller state
995 if (callee_jvms != nullptr && val->is_InlineType() && val->as_InlineType()->is_larval() &&
996 callee_jvms->method()->is_object_constructor() && val == in_map->argument(in_jvms, 0) &&
997 val->bottom_type()->is_inlinetypeptr()) {
998 val = callee_jvms->map()->local(callee_jvms, 0); // Receiver
999 }
1000 call->set_req(p++, val);
1001 }
1002 } else {
1003 p += l; // already set to top above by add_req_batch
1004 }
1005
1006 // Add the Expression Stack
1007 k = in_jvms->stkoff();
1008 l = in_jvms->sp();
1009 out_jvms->set_stkoff(p);
1010 if (!can_prune_locals) {
1011 for (j = 0; j < l; j++) {
1012 Node* val = in_map->in(k + j);
1013 // Check if there's a larval that has been written in the callee state (constructor) and update it in the caller state
1014 if (callee_jvms != nullptr && val->is_InlineType() && val->as_InlineType()->is_larval() &&
1015 callee_jvms->method()->is_object_constructor() && val == in_map->argument(in_jvms, 0) &&
1016 val->bottom_type()->is_inlinetypeptr()) {
1017 val = callee_jvms->map()->local(callee_jvms, 0); // Receiver
1018 }
1019 call->set_req(p++, val);
1020 }
1021 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1022 // Divide stack into {S0,...,S1}, where S0 is set to top.
1023 uint s1 = stack_slots_not_pruned;
1024 stack_slots_not_pruned = 0; // for next iteration
1025 if (s1 > l) s1 = l;
1026 uint s0 = l - s1;
1027 p += s0; // skip the tops preinstalled by add_req_batch
1028 for (j = s0; j < l; j++)
1029 call->set_req(p++, in_map->in(k+j));
1030 } else {
1031 p += l; // already set to top above by add_req_batch
1032 }
1033
1034 // Add the Monitors
1035 k = in_jvms->monoff();
1036 l = in_jvms->mon_size();
1037 out_jvms->set_monoff(p);
1038 for (j = 0; j < l; j++)
1039 call->set_req(p++, in_map->in(k+j));
1040
1041 // Copy any scalar object fields.
1042 k = in_jvms->scloff();
1043 l = in_jvms->scl_size();
1044 out_jvms->set_scloff(p);
1045 for (j = 0; j < l; j++)
1046 call->set_req(p++, in_map->in(k+j));
1047
1048 // Finish the new jvms.
1049 out_jvms->set_endoff(p);
1050
1051 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1052 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1053 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1054 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1055 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1056 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1057
1058 // Update the two tail pointers in parallel.
1059 callee_jvms = out_jvms;
1060 out_jvms = out_jvms->caller();
1061 in_jvms = in_jvms->caller();
1062 }
1063
1064 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1065
1066 // Test the correctness of JVMState::debug_xxx accessors:
1067 assert(call->jvms()->debug_start() == non_debug_edges, "");
1068 assert(call->jvms()->debug_end() == call->req(), "");
1069 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1070 }
1071
1072 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1073 Bytecodes::Code code = java_bc();
1074 if (code == Bytecodes::_wide) {
1075 code = method()->java_code_at_bci(bci() + 1);
1076 }
1077
1078 if (code != Bytecodes::_illegal) {
1079 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1215 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1216 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1217 return _gvn.transform( new AndLNode(conv, mask) );
1218 }
1219
1220 Node* GraphKit::ConvL2I(Node* offset) {
1221 // short-circuit a common case
1222 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1223 if (offset_con != (jlong)Type::OffsetBot) {
1224 return intcon((int) offset_con);
1225 }
1226 return _gvn.transform( new ConvL2INode(offset));
1227 }
1228
1229 //-------------------------load_object_klass-----------------------------------
1230 Node* GraphKit::load_object_klass(Node* obj) {
1231 // Special-case a fresh allocation to avoid building nodes:
1232 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1233 if (akls != nullptr) return akls;
1234 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1235 return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1236 }
1237
1238 //-------------------------load_array_length-----------------------------------
1239 Node* GraphKit::load_array_length(Node* array) {
1240 // Special-case a fresh allocation to avoid building nodes:
1241 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1242 Node *alen;
1243 if (alloc == nullptr) {
1244 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1245 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1246 } else {
1247 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1248 }
1249 return alen;
1250 }
1251
1252 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1253 const TypeOopPtr* oop_type,
1254 bool replace_length_in_map) {
1255 Node* length = alloc->Ideal_length();
1264 replace_in_map(length, ccast);
1265 }
1266 return ccast;
1267 }
1268 }
1269 return length;
1270 }
1271
1272 //------------------------------do_null_check----------------------------------
1273 // Helper function to do a null pointer check. Returned value is
1274 // the incoming address with null casted away. You are allowed to use the
1275 // not-null value only if you are control dependent on the test.
1276 #ifndef PRODUCT
1277 extern uint explicit_null_checks_inserted,
1278 explicit_null_checks_elided;
1279 #endif
1280 Node* GraphKit::null_check_common(Node* value, BasicType type,
1281 // optional arguments for variations:
1282 bool assert_null,
1283 Node* *null_control,
1284 bool speculative,
1285 bool is_init_check) {
1286 assert(!assert_null || null_control == nullptr, "not both at once");
1287 if (stopped()) return top();
1288 NOT_PRODUCT(explicit_null_checks_inserted++);
1289
1290 if (value->is_InlineType()) {
1291 // Null checking a scalarized but nullable inline type. Check the IsInit
1292 // input instead of the oop input to avoid keeping buffer allocations alive.
1293 InlineTypeNode* vtptr = value->as_InlineType();
1294 while (vtptr->get_oop()->is_InlineType()) {
1295 vtptr = vtptr->get_oop()->as_InlineType();
1296 }
1297 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true);
1298 if (stopped()) {
1299 return top();
1300 }
1301 if (assert_null) {
1302 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1303 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1304 // replace_in_map(value, vtptr);
1305 // return vtptr;
1306 replace_in_map(value, null());
1307 return null();
1308 }
1309 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1310 return cast_not_null(value, do_replace_in_map);
1311 }
1312
1313 // Construct null check
1314 Node *chk = nullptr;
1315 switch(type) {
1316 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1317 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1318 case T_ARRAY : // fall through
1319 type = T_OBJECT; // simplify further tests
1320 case T_OBJECT : {
1321 const Type *t = _gvn.type( value );
1322
1323 const TypeOopPtr* tp = t->isa_oopptr();
1324 if (tp != nullptr && !tp->is_loaded()
1325 // Only for do_null_check, not any of its siblings:
1326 && !assert_null && null_control == nullptr) {
1327 // Usually, any field access or invocation on an unloaded oop type
1328 // will simply fail to link, since the statically linked class is
1329 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1330 // the static class is loaded but the sharper oop type is not.
1331 // Rather than checking for this obscure case in lots of places,
1332 // we simply observe that a null check on an unloaded class
1396 }
1397 Node *oldcontrol = control();
1398 set_control(cfg);
1399 Node *res = cast_not_null(value);
1400 set_control(oldcontrol);
1401 NOT_PRODUCT(explicit_null_checks_elided++);
1402 return res;
1403 }
1404 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1405 if (cfg == nullptr) break; // Quit at region nodes
1406 depth++;
1407 }
1408 }
1409
1410 //-----------
1411 // Branch to failure if null
1412 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1413 Deoptimization::DeoptReason reason;
1414 if (assert_null) {
1415 reason = Deoptimization::reason_null_assert(speculative);
1416 } else if (type == T_OBJECT || is_init_check) {
1417 reason = Deoptimization::reason_null_check(speculative);
1418 } else {
1419 reason = Deoptimization::Reason_div0_check;
1420 }
1421 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1422 // ciMethodData::has_trap_at will return a conservative -1 if any
1423 // must-be-null assertion has failed. This could cause performance
1424 // problems for a method after its first do_null_assert failure.
1425 // Consider using 'Reason_class_check' instead?
1426
1427 // To cause an implicit null check, we set the not-null probability
1428 // to the maximum (PROB_MAX). For an explicit check the probability
1429 // is set to a smaller value.
1430 if (null_control != nullptr || too_many_traps(reason)) {
1431 // probability is less likely
1432 ok_prob = PROB_LIKELY_MAG(3);
1433 } else if (!assert_null &&
1434 (ImplicitNullCheckThreshold > 0) &&
1435 method() != nullptr &&
1436 (method()->method_data()->trap_count(reason)
1470 }
1471
1472 if (assert_null) {
1473 // Cast obj to null on this path.
1474 replace_in_map(value, zerocon(type));
1475 return zerocon(type);
1476 }
1477
1478 // Cast obj to not-null on this path, if there is no null_control.
1479 // (If there is a null_control, a non-null value may come back to haunt us.)
1480 if (type == T_OBJECT) {
1481 Node* cast = cast_not_null(value, false);
1482 if (null_control == nullptr || (*null_control) == top())
1483 replace_in_map(value, cast);
1484 value = cast;
1485 }
1486
1487 return value;
1488 }
1489
1490 //------------------------------cast_not_null----------------------------------
1491 // Cast obj to not-null on this path
1492 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1493 if (obj->is_InlineType()) {
1494 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1495 vt->as_InlineType()->set_is_init(_gvn);
1496 vt = _gvn.transform(vt);
1497 if (do_replace_in_map) {
1498 replace_in_map(obj, vt);
1499 }
1500 return vt;
1501 }
1502 const Type *t = _gvn.type(obj);
1503 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1504 // Object is already not-null?
1505 if( t == t_not_null ) return obj;
1506
1507 Node* cast = new CastPPNode(control(), obj,t_not_null);
1508 cast = _gvn.transform( cast );
1509
1510 // Scan for instances of 'obj' in the current JVM mapping.
1511 // These instances are known to be not-null after the test.
1512 if (do_replace_in_map)
1513 replace_in_map(obj, cast);
1514
1515 return cast; // Return casted value
1516 }
1517
1518 // Sometimes in intrinsics, we implicitly know an object is not null
1519 // (there's no actual null check) so we can cast it to not null. In
1520 // the course of optimizations, the input to the cast can become null.
1521 // In that case that data path will die and we need the control path
1604 }
1605
1606 //=============================================================================
1607 //
1608 // parser factory methods for MemNodes
1609 //
1610 // These are layered on top of the factory methods in LoadNode and StoreNode,
1611 // and integrate with the parser's memory state and _gvn engine.
1612 //
1613
1614 // factory methods in "int adr_idx"
1615 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1616 int adr_idx,
1617 MemNode::MemOrd mo,
1618 LoadNode::ControlDependency control_dependency,
1619 bool require_atomic_access,
1620 bool unaligned,
1621 bool mismatched,
1622 bool unsafe,
1623 uint8_t barrier_data) {
1624 // Fix 8344108 and renable the commented assert
1625 //assert(adr_idx == C->get_alias_index(_gvn.type(adr)->isa_ptr()), "slice of address and input slice don't match");
1626 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1627 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1628 debug_only(adr_type = C->get_adr_type(adr_idx));
1629 Node* mem = memory(adr_idx);
1630 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1631 ld = _gvn.transform(ld);
1632
1633 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1634 // Improve graph before escape analysis and boxing elimination.
1635 record_for_igvn(ld);
1636 if (ld->is_DecodeN()) {
1637 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1638 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1639 // a Phi). Recording such cases is still perfectly sound, but may be
1640 // unnecessary and result in some minor IGVN overhead.
1641 record_for_igvn(ld->in(1));
1642 }
1643 }
1644 return ld;
1645 }
1646
1647 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1648 int adr_idx,
1649 MemNode::MemOrd mo,
1650 bool require_atomic_access,
1651 bool unaligned,
1652 bool mismatched,
1653 bool unsafe,
1654 int barrier_data) {
1655 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1656 // Fix 8344108 and renable the commented assert
1657 //assert(adr_idx == C->get_alias_index(_gvn.type(adr)->isa_ptr()), "slice of address and input slice don't match");
1658 const TypePtr* adr_type = nullptr;
1659 debug_only(adr_type = C->get_adr_type(adr_idx));
1660 Node *mem = memory(adr_idx);
1661 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1662 if (unaligned) {
1663 st->as_Store()->set_unaligned_access();
1664 }
1665 if (mismatched) {
1666 st->as_Store()->set_mismatched_access();
1667 }
1668 if (unsafe) {
1669 st->as_Store()->set_unsafe_access();
1670 }
1671 st->as_Store()->set_barrier_data(barrier_data);
1672 st = _gvn.transform(st);
1673 set_memory(st, adr_idx);
1674 // Back-to-back stores can only remove intermediate store with DU info
1675 // so push on worklist for optimizer.
1676 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1677 record_for_igvn(st);
1678
1679 return st;
1680 }
1681
1682 Node* GraphKit::access_store_at(Node* obj,
1683 Node* adr,
1684 const TypePtr* adr_type,
1685 Node* val,
1686 const Type* val_type,
1687 BasicType bt,
1688 DecoratorSet decorators,
1689 bool safe_for_replace) {
1690 // Transformation of a value which could be null pointer (CastPP #null)
1691 // could be delayed during Parse (for example, in adjust_map_after_if()).
1692 // Execute transformation here to avoid barrier generation in such case.
1693 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1694 val = _gvn.makecon(TypePtr::NULL_PTR);
1695 }
1696
1697 if (stopped()) {
1698 return top(); // Dead path ?
1699 }
1700
1701 assert(val != nullptr, "not dead path");
1702 if (val->is_InlineType()) {
1703 // Store to non-flat field. Buffer the inline type and make sure
1704 // the store is re-executed if the allocation triggers deoptimization.
1705 PreserveReexecuteState preexecs(this);
1706 jvms()->set_should_reexecute(true);
1707 val = val->as_InlineType()->buffer(this, safe_for_replace);
1708 }
1709
1710 C2AccessValuePtr addr(adr, adr_type);
1711 C2AccessValue value(val, val_type);
1712 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1713 if (access.is_raw()) {
1714 return _barrier_set->BarrierSetC2::store_at(access, value);
1715 } else {
1716 return _barrier_set->store_at(access, value);
1717 }
1718 }
1719
1720 Node* GraphKit::access_load_at(Node* obj, // containing obj
1721 Node* adr, // actual address to store val at
1722 const TypePtr* adr_type,
1723 const Type* val_type,
1724 BasicType bt,
1725 DecoratorSet decorators,
1726 Node* ctl) {
1727 if (stopped()) {
1728 return top(); // Dead path ?
1729 }
1730
1731 C2AccessValuePtr addr(adr, adr_type);
1732 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1733 if (access.is_raw()) {
1734 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1735 } else {
1736 return _barrier_set->load_at(access, val_type);
1737 }
1738 }
1739
1740 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1741 const Type* val_type,
1742 BasicType bt,
1743 DecoratorSet decorators) {
1744 if (stopped()) {
1745 return top(); // Dead path ?
1746 }
1747
1748 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1749 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1750 if (access.is_raw()) {
1751 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1752 } else {
1817 Node* new_val,
1818 const Type* value_type,
1819 BasicType bt,
1820 DecoratorSet decorators) {
1821 C2AccessValuePtr addr(adr, adr_type);
1822 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1823 if (access.is_raw()) {
1824 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1825 } else {
1826 return _barrier_set->atomic_add_at(access, new_val, value_type);
1827 }
1828 }
1829
1830 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1831 return _barrier_set->clone(this, src, dst, size, is_array);
1832 }
1833
1834 //-------------------------array_element_address-------------------------
1835 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1836 const TypeInt* sizetype, Node* ctrl) {
1837 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1838 uint shift = arytype->is_flat() ? arytype->flat_log_elem_size() : exact_log2(type2aelembytes(elembt));
1839 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1840
1841 // short-circuit a common case (saves lots of confusing waste motion)
1842 jint idx_con = find_int_con(idx, -1);
1843 if (idx_con >= 0) {
1844 intptr_t offset = header + ((intptr_t)idx_con << shift);
1845 return basic_plus_adr(ary, offset);
1846 }
1847
1848 // must be correct type for alignment purposes
1849 Node* base = basic_plus_adr(ary, header);
1850 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1851 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1852 return basic_plus_adr(ary, base, scale);
1853 }
1854
1855 //-------------------------load_array_element-------------------------
1856 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1857 const Type* elemtype = arytype->elem();
1858 BasicType elembt = elemtype->array_element_basic_type();
1859 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1860 if (elembt == T_NARROWOOP) {
1861 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1862 }
1863 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1864 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1865 return ld;
1866 }
1867
1868 //-------------------------set_arguments_for_java_call-------------------------
1869 // Arguments (pre-popped from the stack) are taken from the JVMS.
1870 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1871 PreserveReexecuteState preexecs(this);
1872 if (EnableValhalla) {
1873 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1874 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1875 jvms()->set_should_reexecute(true);
1876 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1877 inc_sp(arg_size);
1878 }
1879 // Add the call arguments
1880 const TypeTuple* domain = call->tf()->domain_sig();
1881 uint nargs = domain->cnt();
1882 int arg_num = 0;
1883 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1884 Node* arg = argument(i-TypeFunc::Parms);
1885 const Type* t = domain->field_at(i);
1886 // TODO 8284443 A static call to a mismatched method should still be scalarized
1887 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1888 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1889 if (!arg->is_InlineType()) {
1890 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1891 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass(), t->inline_klass()->is_null_free());
1892 }
1893 InlineTypeNode* vt = arg->as_InlineType();
1894 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1895 // If an inline type argument is passed as fields, attach the Method* to the call site
1896 // to be able to access the extended signature later via attached_method_before_pc().
1897 // For example, see CompiledMethod::preserve_callee_argument_oops().
1898 call->set_override_symbolic_info(true);
1899 // Register an evol dependency on the callee method to make sure that this method is deoptimized and
1900 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1901 C->dependencies()->assert_evol_method(call->method());
1902 arg_num++;
1903 continue;
1904 } else if (arg->is_InlineType()) {
1905 // Pass inline type argument via oop to callee
1906 InlineTypeNode* inline_type = arg->as_InlineType();
1907 const ciMethod* method = call->method();
1908 ciInstanceKlass* holder = method->holder();
1909 const bool is_receiver = (i == TypeFunc::Parms);
1910 const bool is_abstract_or_object_klass_constructor = method->is_object_constructor() &&
1911 (holder->is_abstract() || holder->is_java_lang_Object());
1912 const bool is_larval_receiver_on_super_constructor = is_receiver && is_abstract_or_object_klass_constructor;
1913 bool must_init_buffer = true;
1914 // We always need to buffer inline types when they are escaping. However, we can skip the actual initialization
1915 // of the buffer if the inline type is a larval because we are going to update the buffer anyway which requires
1916 // us to create a new one. But there is one special case where we are still required to initialize the buffer:
1917 // When we have a larval receiver invoked on an abstract (value class) constructor or the Object constructor (that
1918 // is not going to be inlined). After this call, the larval is completely initialized and thus not a larval anymore.
1919 // We therefore need to force an initialization of the buffer to not lose all the field writes so far in case the
1920 // buffer needs to be used (e.g. to read from when deoptimizing at runtime) or further updated in abstract super
1921 // value class constructors which could have more fields to be initialized. Note that we do not need to
1922 // initialize the buffer when invoking another constructor in the same class on a larval receiver because we
1923 // have not initialized any fields, yet (this is done completely by the other constructor call).
1924 if (inline_type->is_larval() && !is_larval_receiver_on_super_constructor) {
1925 must_init_buffer = false;
1926 }
1927 arg = inline_type->buffer(this, true, must_init_buffer);
1928 }
1929 if (t != Type::HALF) {
1930 arg_num++;
1931 }
1932 call->init_req(idx++, arg);
1933 }
1934 }
1935
1936 //---------------------------set_edges_for_java_call---------------------------
1937 // Connect a newly created call into the current JVMS.
1938 // A return value node (if any) is returned from set_edges_for_java_call.
1939 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1940
1941 // Add the predefined inputs:
1942 call->init_req( TypeFunc::Control, control() );
1943 call->init_req( TypeFunc::I_O , i_o() );
1944 call->init_req( TypeFunc::Memory , reset_memory() );
1945 call->init_req( TypeFunc::FramePtr, frameptr() );
1946 call->init_req( TypeFunc::ReturnAdr, top() );
1947
1948 add_safepoint_edges(call, must_throw);
1949
1950 Node* xcall = _gvn.transform(call);
1951
1952 if (xcall == top()) {
1953 set_control(top());
1954 return;
1955 }
1956 assert(xcall == call, "call identity is stable");
1957
1958 // Re-use the current map to produce the result.
1959
1960 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1961 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1962 set_all_memory_call(xcall, separate_io_proj);
1963
1964 //return xcall; // no need, caller already has it
1965 }
1966
1967 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1968 if (stopped()) return top(); // maybe the call folded up?
1969
1970 // Note: Since any out-of-line call can produce an exception,
1971 // we always insert an I_O projection from the call into the result.
1972
1973 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1974
1975 if (separate_io_proj) {
1976 // The caller requested separate projections be used by the fall
1977 // through and exceptional paths, so replace the projections for
1978 // the fall through path.
1979 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1980 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1981 }
1982
1983 // Capture the return value, if any.
1984 Node* ret;
1985 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
1986 ret = top();
1987 } else if (call->tf()->returns_inline_type_as_fields()) {
1988 // Return of multiple values (inline type fields): we create a
1989 // InlineType node, each field is a projection from the call.
1990 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
1991 uint base_input = TypeFunc::Parms;
1992 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
1993 } else {
1994 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1995 ciType* t = call->method()->return_type();
1996 if (t->is_klass()) {
1997 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
1998 if (type->is_inlinetypeptr()) {
1999 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass(), type->inline_klass()->is_null_free());
2000 }
2001 }
2002 }
2003
2004 // We just called the constructor on a value type receiver. Reload it from the buffer
2005 ciMethod* method = call->method();
2006 if (method->is_object_constructor() && !method->holder()->is_java_lang_Object()) {
2007 InlineTypeNode* inline_type_receiver = call->in(TypeFunc::Parms)->isa_InlineType();
2008 if (inline_type_receiver != nullptr) {
2009 assert(inline_type_receiver->is_larval(), "must be larval");
2010 assert(inline_type_receiver->is_allocated(&gvn()), "larval must be buffered");
2011 InlineTypeNode* reloaded = InlineTypeNode::make_from_oop(this, inline_type_receiver->get_oop(),
2012 inline_type_receiver->bottom_type()->inline_klass(), true);
2013 assert(!reloaded->is_larval(), "should not be larval anymore");
2014 replace_in_map(inline_type_receiver, reloaded);
2015 }
2016 }
2017
2018 return ret;
2019 }
2020
2021 //--------------------set_predefined_input_for_runtime_call--------------------
2022 // Reading and setting the memory state is way conservative here.
2023 // The real problem is that I am not doing real Type analysis on memory,
2024 // so I cannot distinguish card mark stores from other stores. Across a GC
2025 // point the Store Barrier and the card mark memory has to agree. I cannot
2026 // have a card mark store and its barrier split across the GC point from
2027 // either above or below. Here I get that to happen by reading ALL of memory.
2028 // A better answer would be to separate out card marks from other memory.
2029 // For now, return the input memory state, so that it can be reused
2030 // after the call, if this call has restricted memory effects.
2031 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2032 // Set fixed predefined input arguments
2033 Node* memory = reset_memory();
2034 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2035 call->init_req( TypeFunc::Control, control() );
2036 call->init_req( TypeFunc::I_O, top() ); // does no i/o
2037 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
2088 if (use->is_MergeMem()) {
2089 wl.push(use);
2090 }
2091 }
2092 }
2093
2094 // Replace the call with the current state of the kit.
2095 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2096 JVMState* ejvms = nullptr;
2097 if (has_exceptions()) {
2098 ejvms = transfer_exceptions_into_jvms();
2099 }
2100
2101 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2102 ReplacedNodes replaced_nodes_exception;
2103 Node* ex_ctl = top();
2104
2105 SafePointNode* final_state = stop();
2106
2107 // Find all the needed outputs of this call
2108 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2109
2110 Unique_Node_List wl;
2111 Node* init_mem = call->in(TypeFunc::Memory);
2112 Node* final_mem = final_state->in(TypeFunc::Memory);
2113 Node* final_ctl = final_state->in(TypeFunc::Control);
2114 Node* final_io = final_state->in(TypeFunc::I_O);
2115
2116 // Replace all the old call edges with the edges from the inlining result
2117 if (callprojs->fallthrough_catchproj != nullptr) {
2118 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2119 }
2120 if (callprojs->fallthrough_memproj != nullptr) {
2121 if (final_mem->is_MergeMem()) {
2122 // Parser's exits MergeMem was not transformed but may be optimized
2123 final_mem = _gvn.transform(final_mem);
2124 }
2125 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2126 add_mergemem_users_to_worklist(wl, final_mem);
2127 }
2128 if (callprojs->fallthrough_ioproj != nullptr) {
2129 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2130 }
2131
2132 // Replace the result with the new result if it exists and is used
2133 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2134 // If the inlined code is dead, the result projections for an inline type returned as
2135 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2136 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2137 "unexpected number of results");
2138 C->gvn_replace_by(callprojs->resproj[0], result);
2139 }
2140
2141 if (ejvms == nullptr) {
2142 // No exception edges to simply kill off those paths
2143 if (callprojs->catchall_catchproj != nullptr) {
2144 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2145 }
2146 if (callprojs->catchall_memproj != nullptr) {
2147 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2148 }
2149 if (callprojs->catchall_ioproj != nullptr) {
2150 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2151 }
2152 // Replace the old exception object with top
2153 if (callprojs->exobj != nullptr) {
2154 C->gvn_replace_by(callprojs->exobj, C->top());
2155 }
2156 } else {
2157 GraphKit ekit(ejvms);
2158
2159 // Load my combined exception state into the kit, with all phis transformed:
2160 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2161 replaced_nodes_exception = ex_map->replaced_nodes();
2162
2163 Node* ex_oop = ekit.use_exception_state(ex_map);
2164
2165 if (callprojs->catchall_catchproj != nullptr) {
2166 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2167 ex_ctl = ekit.control();
2168 }
2169 if (callprojs->catchall_memproj != nullptr) {
2170 Node* ex_mem = ekit.reset_memory();
2171 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2172 add_mergemem_users_to_worklist(wl, ex_mem);
2173 }
2174 if (callprojs->catchall_ioproj != nullptr) {
2175 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2176 }
2177
2178 // Replace the old exception object with the newly created one
2179 if (callprojs->exobj != nullptr) {
2180 C->gvn_replace_by(callprojs->exobj, ex_oop);
2181 }
2182 }
2183
2184 // Disconnect the call from the graph
2185 call->disconnect_inputs(C);
2186 C->gvn_replace_by(call, C->top());
2187
2188 // Clean up any MergeMems that feed other MergeMems since the
2189 // optimizer doesn't like that.
2190 while (wl.size() > 0) {
2191 _gvn.transform(wl.pop());
2192 }
2193
2194 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2195 replaced_nodes.apply(C, final_ctl);
2196 }
2197 if (!ex_ctl->is_top() && do_replaced_nodes) {
2198 replaced_nodes_exception.apply(C, ex_ctl);
2199 }
2200 }
2201
2202
2203 //------------------------------increment_counter------------------------------
2204 // for statistics: increment a VM counter by 1
2205
2206 void GraphKit::increment_counter(address counter_addr) {
2207 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2208 increment_counter(adr1);
2209 }
2210
2211 void GraphKit::increment_counter(Node* counter_addr) {
2212 int adr_type = Compile::AliasIdxRaw;
2213 Node* ctrl = control();
2214 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2375 *
2376 * @param n node that the type applies to
2377 * @param exact_kls type from profiling
2378 * @param maybe_null did profiling see null?
2379 *
2380 * @return node with improved type
2381 */
2382 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2383 const Type* current_type = _gvn.type(n);
2384 assert(UseTypeSpeculation, "type speculation must be on");
2385
2386 const TypePtr* speculative = current_type->speculative();
2387
2388 // Should the klass from the profile be recorded in the speculative type?
2389 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2390 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2391 const TypeOopPtr* xtype = tklass->as_instance_type();
2392 assert(xtype->klass_is_exact(), "Should be exact");
2393 // Any reason to believe n is not null (from this profiling or a previous one)?
2394 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2395 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2396 // record the new speculative type's depth
2397 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2398 speculative = speculative->with_inline_depth(jvms()->depth());
2399 } else if (current_type->would_improve_ptr(ptr_kind)) {
2400 // Profiling report that null was never seen so we can change the
2401 // speculative type to non null ptr.
2402 if (ptr_kind == ProfileAlwaysNull) {
2403 speculative = TypePtr::NULL_PTR;
2404 } else {
2405 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2406 const TypePtr* ptr = TypePtr::NOTNULL;
2407 if (speculative != nullptr) {
2408 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2409 } else {
2410 speculative = ptr;
2411 }
2412 }
2413 }
2414
2415 if (speculative != current_type->speculative()) {
2416 // Build a type with a speculative type (what we think we know
2417 // about the type but will need a guard when we use it)
2418 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2419 // We're changing the type, we need a new CheckCast node to carry
2420 // the new type. The new type depends on the control: what
2421 // profiling tells us is only valid from here as far as we can
2422 // tell.
2423 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2424 cast = _gvn.transform(cast);
2425 replace_in_map(n, cast);
2426 n = cast;
2427 }
2428
2429 return n;
2430 }
2431
2432 /**
2433 * Record profiling data from receiver profiling at an invoke with the
2434 * type system so that it can propagate it (speculation)
2435 *
2436 * @param n receiver node
2437 *
2438 * @return node with improved type
2439 */
2440 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2441 if (!UseTypeSpeculation) {
2442 return n;
2443 }
2444 ciKlass* exact_kls = profile_has_unique_klass();
2445 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2446 if ((java_bc() == Bytecodes::_checkcast ||
2447 java_bc() == Bytecodes::_instanceof ||
2448 java_bc() == Bytecodes::_aastore) &&
2449 method()->method_data()->is_mature()) {
2450 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2451 if (data != nullptr) {
2452 if (java_bc() == Bytecodes::_aastore) {
2453 ciKlass* array_type = nullptr;
2454 ciKlass* element_type = nullptr;
2455 ProfilePtrKind element_ptr = ProfileMaybeNull;
2456 bool flat_array = true;
2457 bool null_free_array = true;
2458 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2459 exact_kls = element_type;
2460 ptr_kind = element_ptr;
2461 } else {
2462 if (!data->as_BitData()->null_seen()) {
2463 ptr_kind = ProfileNeverNull;
2464 } else {
2465 assert(data->is_ReceiverTypeData(), "bad profile data type");
2466 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2467 uint i = 0;
2468 for (; i < call->row_limit(); i++) {
2469 ciKlass* receiver = call->receiver(i);
2470 if (receiver != nullptr) {
2471 break;
2472 }
2473 }
2474 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2475 }
2476 }
2477 }
2478 }
2479 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2480 }
2481
2482 /**
2483 * Record profiling data from argument profiling at an invoke with the
2484 * type system so that it can propagate it (speculation)
2485 *
2486 * @param dest_method target method for the call
2487 * @param bc what invoke bytecode is this?
2488 */
2489 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2490 if (!UseTypeSpeculation) {
2491 return;
2492 }
2493 const TypeFunc* tf = TypeFunc::make(dest_method);
2494 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2495 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2496 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2497 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2498 if (is_reference_type(targ->basic_type())) {
2499 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2500 ciKlass* better_type = nullptr;
2501 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2502 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2503 }
2504 i++;
2505 }
2506 }
2507 }
2508
2509 /**
2510 * Record profiling data from parameter profiling at an invoke with
2511 * the type system so that it can propagate it (speculation)
2512 */
2513 void GraphKit::record_profiled_parameters_for_speculation() {
2514 if (!UseTypeSpeculation) {
2515 return;
2516 }
2517 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2531 * the type system so that it can propagate it (speculation)
2532 */
2533 void GraphKit::record_profiled_return_for_speculation() {
2534 if (!UseTypeSpeculation) {
2535 return;
2536 }
2537 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2538 ciKlass* better_type = nullptr;
2539 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2540 // If profiling reports a single type for the return value,
2541 // feed it to the type system so it can propagate it as a
2542 // speculative type
2543 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2544 }
2545 }
2546
2547 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2548 if (Matcher::strict_fp_requires_explicit_rounding) {
2549 // (Note: TypeFunc::make has a cache that makes this fast.)
2550 const TypeFunc* tf = TypeFunc::make(dest_method);
2551 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2552 for (int j = 0; j < nargs; j++) {
2553 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2554 if (targ->basic_type() == T_DOUBLE) {
2555 // If any parameters are doubles, they must be rounded before
2556 // the call, dprecision_rounding does gvn.transform
2557 Node *arg = argument(j);
2558 arg = dprecision_rounding(arg);
2559 set_argument(j, arg);
2560 }
2561 }
2562 }
2563 }
2564
2565 // rounding for strict float precision conformance
2566 Node* GraphKit::precision_rounding(Node* n) {
2567 if (Matcher::strict_fp_requires_explicit_rounding) {
2568 #ifdef IA32
2569 if (UseSSE == 0) {
2570 return _gvn.transform(new RoundFloatNode(nullptr, n));
2571 }
2572 #else
2573 Unimplemented();
2682 // The first null ends the list.
2683 Node* parm0, Node* parm1,
2684 Node* parm2, Node* parm3,
2685 Node* parm4, Node* parm5,
2686 Node* parm6, Node* parm7) {
2687 assert(call_addr != nullptr, "must not call null targets");
2688
2689 // Slow-path call
2690 bool is_leaf = !(flags & RC_NO_LEAF);
2691 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2692 if (call_name == nullptr) {
2693 assert(!is_leaf, "must supply name for leaf");
2694 call_name = OptoRuntime::stub_name(call_addr);
2695 }
2696 CallNode* call;
2697 if (!is_leaf) {
2698 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2699 } else if (flags & RC_NO_FP) {
2700 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2701 } else if (flags & RC_VECTOR){
2702 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2703 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2704 } else {
2705 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2706 }
2707
2708 // The following is similar to set_edges_for_java_call,
2709 // except that the memory effects of the call are restricted to AliasIdxRaw.
2710
2711 // Slow path call has no side-effects, uses few values
2712 bool wide_in = !(flags & RC_NARROW_MEM);
2713 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2714
2715 Node* prev_mem = nullptr;
2716 if (wide_in) {
2717 prev_mem = set_predefined_input_for_runtime_call(call);
2718 } else {
2719 assert(!wide_out, "narrow in => narrow out");
2720 Node* narrow_mem = memory(adr_type);
2721 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2722 }
2762
2763 if (has_io) {
2764 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2765 }
2766 return call;
2767
2768 }
2769
2770 // i2b
2771 Node* GraphKit::sign_extend_byte(Node* in) {
2772 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2773 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2774 }
2775
2776 // i2s
2777 Node* GraphKit::sign_extend_short(Node* in) {
2778 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2779 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2780 }
2781
2782
2783 //------------------------------merge_memory-----------------------------------
2784 // Merge memory from one path into the current memory state.
2785 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2786 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2787 Node* old_slice = mms.force_memory();
2788 Node* new_slice = mms.memory2();
2789 if (old_slice != new_slice) {
2790 PhiNode* phi;
2791 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2792 if (mms.is_empty()) {
2793 // clone base memory Phi's inputs for this memory slice
2794 assert(old_slice == mms.base_memory(), "sanity");
2795 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2796 _gvn.set_type(phi, Type::MEMORY);
2797 for (uint i = 1; i < phi->req(); i++) {
2798 phi->init_req(i, old_slice->in(i));
2799 }
2800 } else {
2801 phi = old_slice->as_Phi(); // Phi was generated already
2802 }
3065
3066 // Now do a linear scan of the secondary super-klass array. Again, no real
3067 // performance impact (too rare) but it's gotta be done.
3068 // Since the code is rarely used, there is no penalty for moving it
3069 // out of line, and it can only improve I-cache density.
3070 // The decision to inline or out-of-line this final check is platform
3071 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3072 Node* psc = gvn.transform(
3073 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3074
3075 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3076 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3077 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3078
3079 // Return false path; set default control to true path.
3080 *ctrl = gvn.transform(r_ok_subtype);
3081 return gvn.transform(r_not_subtype);
3082 }
3083
3084 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3085 const Type* sub_t = _gvn.type(obj_or_subklass);
3086 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3087 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3088 obj_or_subklass = makecon(sub_t);
3089 }
3090 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3091 if (expand_subtype_check) {
3092 MergeMemNode* mem = merged_memory();
3093 Node* ctrl = control();
3094 Node* subklass = obj_or_subklass;
3095 if (!sub_t->isa_klassptr()) {
3096 subklass = load_object_klass(obj_or_subklass);
3097 }
3098
3099 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3100 set_control(ctrl);
3101 return n;
3102 }
3103
3104 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3105 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3106 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3107 set_control(_gvn.transform(new IfTrueNode(iff)));
3108 return _gvn.transform(new IfFalseNode(iff));
3109 }
3110
3111 // Profile-driven exact type check:
3112 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3113 float prob, Node* *casted_receiver) {
3114 assert(!klass->is_interface(), "no exact type check on interfaces");
3115 Node* fail = top();
3116 const Type* rec_t = _gvn.type(receiver);
3117 if (rec_t->is_inlinetypeptr()) {
3118 if (klass->equals(rec_t->inline_klass())) {
3119 (*casted_receiver) = receiver; // Always passes
3120 } else {
3121 (*casted_receiver) = top(); // Always fails
3122 fail = control();
3123 set_control(top());
3124 }
3125 return fail;
3126 }
3127 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3128 Node* recv_klass = load_object_klass(receiver);
3129 fail = type_check(recv_klass, tklass, prob);
3130
3131 if (!stopped()) {
3132 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3133 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3134 assert(recv_xtype->klass_is_exact(), "");
3135
3136 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3137 // Subsume downstream occurrences of receiver with a cast to
3138 // recv_xtype, since now we know what the type will be.
3139 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3140 Node* res = _gvn.transform(cast);
3141 if (recv_xtype->is_inlinetypeptr()) {
3142 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3143 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3144 }
3145 (*casted_receiver) = res;
3146 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3147 // (User must make the replace_in_map call.)
3148 }
3149 }
3150
3151 return fail;
3152 }
3153
3154 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3155 float prob) {
3156 Node* want_klass = makecon(tklass);
3157 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3158 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3159 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3160 set_control(_gvn.transform(new IfTrueNode (iff)));
3161 Node* fail = _gvn.transform(new IfFalseNode(iff));
3162 return fail;
3163 }
3164
3165 //------------------------------subtype_check_receiver-------------------------
3166 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3167 Node** casted_receiver) {
3168 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3169 Node* want_klass = makecon(tklass);
3170
3171 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3172
3173 // Ignore interface type information until interface types are properly tracked.
3174 if (!stopped() && !klass->is_interface()) {
3175 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3176 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3177 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3178 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3179 if (recv_type->is_inlinetypeptr()) {
3180 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3181 }
3182 (*casted_receiver) = cast;
3183 }
3184 }
3185
3186 return slow_ctl;
3187 }
3188
3189 //------------------------------seems_never_null-------------------------------
3190 // Use null_seen information if it is available from the profile.
3191 // If we see an unexpected null at a type check we record it and force a
3192 // recompile; the offending check will be recompiled to handle nulls.
3193 // If we see several offending BCIs, then all checks in the
3194 // method will be recompiled.
3195 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3196 speculating = !_gvn.type(obj)->speculative_maybe_null();
3197 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3198 if (UncommonNullCast // Cutout for this technique
3199 && obj != null() // And not the -Xcomp stupid case?
3200 && !too_many_traps(reason)
3201 ) {
3202 if (speculating) {
3271
3272 //------------------------maybe_cast_profiled_receiver-------------------------
3273 // If the profile has seen exactly one type, narrow to exactly that type.
3274 // Subsequent type checks will always fold up.
3275 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3276 const TypeKlassPtr* require_klass,
3277 ciKlass* spec_klass,
3278 bool safe_for_replace) {
3279 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3280
3281 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3282
3283 // Make sure we haven't already deoptimized from this tactic.
3284 if (too_many_traps_or_recompiles(reason))
3285 return nullptr;
3286
3287 // (No, this isn't a call, but it's enough like a virtual call
3288 // to use the same ciMethod accessor to get the profile info...)
3289 // If we have a speculative type use it instead of profiling (which
3290 // may not help us)
3291 ciKlass* exact_kls = spec_klass;
3292 if (exact_kls == nullptr) {
3293 if (java_bc() == Bytecodes::_aastore) {
3294 ciKlass* array_type = nullptr;
3295 ciKlass* element_type = nullptr;
3296 ProfilePtrKind element_ptr = ProfileMaybeNull;
3297 bool flat_array = true;
3298 bool null_free_array = true;
3299 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3300 exact_kls = element_type;
3301 } else {
3302 exact_kls = profile_has_unique_klass();
3303 }
3304 }
3305 if (exact_kls != nullptr) {// no cast failures here
3306 if (require_klass == nullptr ||
3307 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3308 // If we narrow the type to match what the type profile sees or
3309 // the speculative type, we can then remove the rest of the
3310 // cast.
3311 // This is a win, even if the exact_kls is very specific,
3312 // because downstream operations, such as method calls,
3313 // will often benefit from the sharper type.
3314 Node* exact_obj = not_null_obj; // will get updated in place...
3315 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3316 &exact_obj);
3317 { PreserveJVMState pjvms(this);
3318 set_control(slow_ctl);
3319 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3320 }
3321 if (safe_for_replace) {
3322 replace_in_map(not_null_obj, exact_obj);
3323 }
3324 return exact_obj;
3414 // If not_null_obj is dead, only null-path is taken
3415 if (stopped()) { // Doing instance-of on a null?
3416 set_control(null_ctl);
3417 return intcon(0);
3418 }
3419 region->init_req(_null_path, null_ctl);
3420 phi ->init_req(_null_path, intcon(0)); // Set null path value
3421 if (null_ctl == top()) {
3422 // Do this eagerly, so that pattern matches like is_diamond_phi
3423 // will work even during parsing.
3424 assert(_null_path == PATH_LIMIT-1, "delete last");
3425 region->del_req(_null_path);
3426 phi ->del_req(_null_path);
3427 }
3428
3429 // Do we know the type check always succeed?
3430 bool known_statically = false;
3431 if (_gvn.type(superklass)->singleton()) {
3432 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3433 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3434 if (subk != nullptr && subk->is_loaded()) {
3435 int static_res = C->static_subtype_check(superk, subk);
3436 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3437 }
3438 }
3439
3440 if (!known_statically) {
3441 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3442 // We may not have profiling here or it may not help us. If we
3443 // have a speculative type use it to perform an exact cast.
3444 ciKlass* spec_obj_type = obj_type->speculative_type();
3445 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3446 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3447 if (stopped()) { // Profile disagrees with this path.
3448 set_control(null_ctl); // Null is the only remaining possibility.
3449 return intcon(0);
3450 }
3451 if (cast_obj != nullptr) {
3452 not_null_obj = cast_obj;
3453 }
3454 }
3470 record_for_igvn(region);
3471
3472 // If we know the type check always succeeds then we don't use the
3473 // profiling data at this bytecode. Don't lose it, feed it to the
3474 // type system as a speculative type.
3475 if (safe_for_replace) {
3476 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3477 replace_in_map(obj, casted_obj);
3478 }
3479
3480 return _gvn.transform(phi);
3481 }
3482
3483 //-------------------------------gen_checkcast---------------------------------
3484 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3485 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3486 // uncommon-trap paths work. Adjust stack after this call.
3487 // If failure_control is supplied and not null, it is filled in with
3488 // the control edge for the cast failure. Otherwise, an appropriate
3489 // uncommon trap or exception is thrown.
3490 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) {
3491 kill_dead_locals(); // Benefit all the uncommon traps
3492 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3493 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3494 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3495 bool safe_for_replace = (failure_control == nullptr);
3496 assert(!null_free || toop->is_inlinetypeptr(), "must be an inline type pointer");
3497
3498 // Fast cutout: Check the case that the cast is vacuously true.
3499 // This detects the common cases where the test will short-circuit
3500 // away completely. We do this before we perform the null check,
3501 // because if the test is going to turn into zero code, we don't
3502 // want a residual null check left around. (Causes a slowdown,
3503 // for example, in some objArray manipulations, such as a[i]=a[j].)
3504 if (improved_klass_ptr_type->singleton()) {
3505 const TypeKlassPtr* kptr = nullptr;
3506 const Type* t = _gvn.type(obj);
3507 if (t->isa_oop_ptr()) {
3508 kptr = t->is_oopptr()->as_klass_type();
3509 } else if (obj->is_InlineType()) {
3510 ciInlineKlass* vk = t->inline_klass();
3511 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3512 }
3513 if (kptr != nullptr) {
3514 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3515 case Compile::SSC_always_true:
3516 // If we know the type check always succeed then we don't use
3517 // the profiling data at this bytecode. Don't lose it, feed it
3518 // to the type system as a speculative type.
3519 obj = record_profiled_receiver_for_speculation(obj);
3520 if (null_free) {
3521 assert(safe_for_replace, "must be");
3522 obj = null_check(obj);
3523 }
3524 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3525 return obj;
3526 case Compile::SSC_always_false:
3527 if (null_free) {
3528 assert(safe_for_replace, "must be");
3529 obj = null_check(obj);
3530 }
3531 // It needs a null check because a null will *pass* the cast check.
3532 if (t->isa_oopptr() != nullptr && !t->is_oopptr()->maybe_null()) {
3533 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3534 Deoptimization::DeoptReason reason = is_aastore ?
3535 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3536 builtin_throw(reason);
3537 return top();
3538 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3539 return null_assert(obj);
3540 }
3541 break; // Fall through to full check
3542 default:
3543 break;
3544 }
3545 }
3546 }
3547
3548 ciProfileData* data = nullptr;
3549 if (failure_control == nullptr) { // use MDO in regular case only
3550 assert(java_bc() == Bytecodes::_aastore ||
3551 java_bc() == Bytecodes::_checkcast,
3552 "interpreter profiles type checks only for these BCs");
3553 if (method()->method_data()->is_mature()) {
3554 data = method()->method_data()->bci_to_data(bci());
3555 }
3556 }
3557
3558 // Make the merge point
3559 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3560 RegionNode* region = new RegionNode(PATH_LIMIT);
3561 Node* phi = new PhiNode(region, toop);
3562 _gvn.set_type(region, Type::CONTROL);
3563 _gvn.set_type(phi, toop);
3564
3565 C->set_has_split_ifs(true); // Has chance for split-if optimization
3566
3567 // Use null-cast information if it is available
3568 bool speculative_not_null = false;
3569 bool never_see_null = ((failure_control == nullptr) // regular case only
3570 && seems_never_null(obj, data, speculative_not_null));
3571
3572 if (obj->is_InlineType()) {
3573 // Re-execute if buffering during triggers deoptimization
3574 PreserveReexecuteState preexecs(this);
3575 jvms()->set_should_reexecute(true);
3576 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3577 }
3578
3579 // Null check; get casted pointer; set region slot 3
3580 Node* null_ctl = top();
3581 Node* not_null_obj = nullptr;
3582 if (null_free) {
3583 assert(safe_for_replace, "must be");
3584 not_null_obj = null_check(obj);
3585 } else {
3586 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3587 }
3588
3589 // If not_null_obj is dead, only null-path is taken
3590 if (stopped()) { // Doing instance-of on a null?
3591 set_control(null_ctl);
3592 if (toop->is_inlinetypeptr()) {
3593 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3594 }
3595 return null();
3596 }
3597 region->init_req(_null_path, null_ctl);
3598 phi ->init_req(_null_path, null()); // Set null path value
3599 if (null_ctl == top()) {
3600 // Do this eagerly, so that pattern matches like is_diamond_phi
3601 // will work even during parsing.
3602 assert(_null_path == PATH_LIMIT-1, "delete last");
3603 region->del_req(_null_path);
3604 phi ->del_req(_null_path);
3605 }
3606
3607 Node* cast_obj = nullptr;
3608 if (improved_klass_ptr_type->klass_is_exact()) {
3609 // The following optimization tries to statically cast the speculative type of the object
3610 // (for example obtained during profiling) to the type of the superklass and then do a
3611 // dynamic check that the type of the object is what we expect. To work correctly
3612 // for checkcast and aastore the type of superklass should be exact.
3613 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3614 // We may not have profiling here or it may not help us. If we have
3615 // a speculative type use it to perform an exact cast.
3616 ciKlass* spec_obj_type = obj_type->speculative_type();
3617 if (spec_obj_type != nullptr || data != nullptr) {
3618 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3619 if (cast_obj != nullptr) {
3620 if (failure_control != nullptr) // failure is now impossible
3621 (*failure_control) = top();
3622 // adjust the type of the phi to the exact klass:
3623 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3624 }
3625 }
3626 }
3627
3628 if (cast_obj == nullptr) {
3629 // Generate the subtype check
3630 Node* improved_superklass = superklass;
3631 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3632 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3633 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3634 // Additionally, the benefit would only be minor in non-constant cases.
3635 improved_superklass = makecon(improved_klass_ptr_type);
3636 }
3637 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3638 // Plug in success path into the merge
3639 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3640 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3641 if (failure_control == nullptr) {
3642 if (not_subtype_ctrl != top()) { // If failure is possible
3643 PreserveJVMState pjvms(this);
3644 set_control(not_subtype_ctrl);
3645 Node* obj_klass = nullptr;
3646 if (not_null_obj->is_InlineType()) {
3647 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3648 } else {
3649 obj_klass = load_object_klass(not_null_obj);
3650 }
3651 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3652 Deoptimization::DeoptReason reason = is_aastore ?
3653 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3654 builtin_throw(reason);
3655 }
3656 } else {
3657 (*failure_control) = not_subtype_ctrl;
3658 }
3659 }
3660
3661 region->init_req(_obj_path, control());
3662 phi ->init_req(_obj_path, cast_obj);
3663
3664 // A merge of null or Casted-NotNull obj
3665 Node* res = _gvn.transform(phi);
3666
3667 // Note I do NOT always 'replace_in_map(obj,result)' here.
3668 // if( tk->klass()->can_be_primary_super() )
3669 // This means that if I successfully store an Object into an array-of-String
3670 // I 'forget' that the Object is really now known to be a String. I have to
3671 // do this because we don't have true union types for interfaces - if I store
3672 // a Baz into an array-of-Interface and then tell the optimizer it's an
3673 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3674 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3675 // replace_in_map( obj, res );
3676
3677 // Return final merged results
3678 set_control( _gvn.transform(region) );
3679 record_for_igvn(region);
3680
3681 bool not_inline = !toop->can_be_inline_type();
3682 bool not_flat_in_array = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flat_in_array());
3683 if (EnableValhalla && not_flat_in_array) {
3684 // Check if obj has been loaded from an array
3685 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3686 Node* array = nullptr;
3687 if (obj->isa_Load()) {
3688 Node* address = obj->in(MemNode::Address);
3689 if (address->isa_AddP()) {
3690 array = address->as_AddP()->in(AddPNode::Base);
3691 }
3692 } else if (obj->is_Phi()) {
3693 Node* region = obj->in(0);
3694 // TODO make this more robust (see JDK-8231346)
3695 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3696 IfNode* iff = region->in(2)->in(0)->isa_If();
3697 if (iff != nullptr) {
3698 iff->is_flat_array_check(&_gvn, &array);
3699 }
3700 }
3701 }
3702 if (array != nullptr) {
3703 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3704 if (ary_t != nullptr && !ary_t->is_flat()) {
3705 if (!ary_t->is_not_null_free() && not_inline) {
3706 // Casting array element to a non-inline-type, mark array as not null-free.
3707 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3708 replace_in_map(array, cast);
3709 } else if (!ary_t->is_not_flat()) {
3710 // Casting array element to a non-flat type, mark array as not flat.
3711 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3712 replace_in_map(array, cast);
3713 }
3714 }
3715 }
3716 }
3717
3718 if (!stopped() && !res->is_InlineType()) {
3719 res = record_profiled_receiver_for_speculation(res);
3720 if (toop->is_inlinetypeptr()) {
3721 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null());
3722 res = vt;
3723 if (safe_for_replace) {
3724 replace_in_map(obj, vt);
3725 replace_in_map(not_null_obj, vt);
3726 replace_in_map(res, vt);
3727 }
3728 }
3729 }
3730 return res;
3731 }
3732
3733 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3734 // Load markword
3735 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3736 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3737 if (check_lock) {
3738 // Check if obj is locked
3739 Node* locked_bit = MakeConX(markWord::unlocked_value);
3740 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3741 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3742 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3743 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3744 _gvn.transform(iff);
3745 Node* locked_region = new RegionNode(3);
3746 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3747
3748 // Unlocked: Use bits from mark word
3749 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3750 mark_phi->init_req(1, mark);
3751
3752 // Locked: Load prototype header from klass
3753 set_control(_gvn.transform(new IfFalseNode(iff)));
3754 // Make loads control dependent to make sure they are only executed if array is locked
3755 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3756 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, control(), C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3757 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
3758 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3759
3760 locked_region->init_req(2, control());
3761 mark_phi->init_req(2, proto);
3762 set_control(_gvn.transform(locked_region));
3763 record_for_igvn(locked_region);
3764
3765 mark = mark_phi;
3766 }
3767
3768 // Now check if mark word bits are set
3769 Node* mask = MakeConX(mask_val);
3770 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3771 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3772 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3773 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3774 }
3775
3776 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3777 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3778 }
3779
3780 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3781 // We can't use immutable memory here because the mark word is mutable.
3782 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3783 // check is moved out of loops (mainly to enable loop unswitching).
3784 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3785 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3786 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3787 }
3788
3789 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3790 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3791 }
3792
3793 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3794 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3795 RegionNode* region = new RegionNode(3);
3796 Node* null_ctl = top();
3797 null_check_oop(val, &null_ctl);
3798 if (null_ctl != top()) {
3799 PreserveJVMState pjvms(this);
3800 set_control(null_ctl);
3801 {
3802 // Deoptimize if null-free array
3803 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3804 inc_sp(nargs);
3805 uncommon_trap(Deoptimization::Reason_null_check,
3806 Deoptimization::Action_none);
3807 }
3808 region->init_req(1, control());
3809 }
3810 region->init_req(2, control());
3811 set_control(_gvn.transform(region));
3812 record_for_igvn(region);
3813 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3814 // Since we were just successfully storing null, the array can't be null free.
3815 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3816 ary_t = ary_t->cast_to_not_null_free();
3817 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3818 if (safe_for_replace) {
3819 replace_in_map(ary, cast);
3820 }
3821 ary = cast;
3822 }
3823 return ary;
3824 }
3825
3826 //------------------------------next_monitor-----------------------------------
3827 // What number should be given to the next monitor?
3828 int GraphKit::next_monitor() {
3829 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3830 int next = current + C->sync_stack_slots();
3831 // Keep the toplevel high water mark current:
3832 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3833 return current;
3834 }
3835
3836 //------------------------------insert_mem_bar---------------------------------
3837 // Memory barrier to avoid floating things around
3838 // The membar serves as a pinch point between both control and all memory slices.
3839 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3840 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3841 mb->init_req(TypeFunc::Control, control());
3842 mb->init_req(TypeFunc::Memory, reset_memory());
3843 Node* membar = _gvn.transform(mb);
3871 }
3872 Node* membar = _gvn.transform(mb);
3873 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3874 if (alias_idx == Compile::AliasIdxBot) {
3875 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3876 } else {
3877 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3878 }
3879 return membar;
3880 }
3881
3882 //------------------------------shared_lock------------------------------------
3883 // Emit locking code.
3884 FastLockNode* GraphKit::shared_lock(Node* obj) {
3885 // bci is either a monitorenter bc or InvocationEntryBci
3886 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3887 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3888
3889 if( !GenerateSynchronizationCode )
3890 return nullptr; // Not locking things?
3891
3892 if (stopped()) // Dead monitor?
3893 return nullptr;
3894
3895 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3896
3897 // Box the stack location
3898 Node* box = new BoxLockNode(next_monitor());
3899 // Check for bailout after new BoxLockNode
3900 if (failing()) { return nullptr; }
3901 box = _gvn.transform(box);
3902 Node* mem = reset_memory();
3903
3904 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3905
3906 // Add monitor to debug info for the slow path. If we block inside the
3907 // slow path and de-opt, we need the monitor hanging around
3908 map()->push_monitor( flock );
3909
3910 const TypeFunc *tf = LockNode::lock_type();
3911 LockNode *lock = new LockNode(C, tf);
3940 }
3941 #endif
3942
3943 return flock;
3944 }
3945
3946
3947 //------------------------------shared_unlock----------------------------------
3948 // Emit unlocking code.
3949 void GraphKit::shared_unlock(Node* box, Node* obj) {
3950 // bci is either a monitorenter bc or InvocationEntryBci
3951 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3952 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3953
3954 if( !GenerateSynchronizationCode )
3955 return;
3956 if (stopped()) { // Dead monitor?
3957 map()->pop_monitor(); // Kill monitor from debug info
3958 return;
3959 }
3960 assert(!obj->is_InlineType(), "should not unlock on inline type");
3961
3962 // Memory barrier to avoid floating things down past the locked region
3963 insert_mem_bar(Op_MemBarReleaseLock);
3964
3965 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3966 UnlockNode *unlock = new UnlockNode(C, tf);
3967 #ifdef ASSERT
3968 unlock->set_dbg_jvms(sync_jvms());
3969 #endif
3970 uint raw_idx = Compile::AliasIdxRaw;
3971 unlock->init_req( TypeFunc::Control, control() );
3972 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3973 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3974 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3975 unlock->init_req( TypeFunc::ReturnAdr, top() );
3976
3977 unlock->init_req(TypeFunc::Parms + 0, obj);
3978 unlock->init_req(TypeFunc::Parms + 1, box);
3979 unlock = _gvn.transform(unlock)->as_Unlock();
3980
3981 Node* mem = reset_memory();
3982
3983 // unlock has no side-effects, sets few values
3984 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3985
3986 // Kill monitor from debug info
3987 map()->pop_monitor( );
3988 }
3989
3990 //-------------------------------get_layout_helper-----------------------------
3991 // If the given klass is a constant or known to be an array,
3992 // fetch the constant layout helper value into constant_value
3993 // and return null. Otherwise, load the non-constant
3994 // layout helper value, and return the node which represents it.
3995 // This two-faced routine is useful because allocation sites
3996 // almost always feature constant types.
3997 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3998 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3999 if (!StressReflectiveCode && klass_t != nullptr) {
4000 bool xklass = klass_t->klass_is_exact();
4001 bool can_be_flat = false;
4002 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4003 if (UseFlatArray && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4004 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4005 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4006 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->flat_in_array());
4007 }
4008 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4009 jint lhelper;
4010 if (klass_t->is_flat()) {
4011 lhelper = ary_type->flat_layout_helper();
4012 } else if (klass_t->isa_aryklassptr()) {
4013 BasicType elem = ary_type->elem()->array_element_basic_type();
4014 if (is_reference_type(elem, true)) {
4015 elem = T_OBJECT;
4016 }
4017 lhelper = Klass::array_layout_helper(elem);
4018 } else {
4019 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4020 }
4021 if (lhelper != Klass::_lh_neutral_value) {
4022 constant_value = lhelper;
4023 return (Node*) nullptr;
4024 }
4025 }
4026 }
4027 constant_value = Klass::_lh_neutral_value; // put in a known value
4028 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
4029 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4030 }
4031
4032 // We just put in an allocate/initialize with a big raw-memory effect.
4033 // Hook selected additional alias categories on the initialization.
4034 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4035 MergeMemNode* init_in_merge,
4036 Node* init_out_raw) {
4037 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4038 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4039
4040 Node* prevmem = kit.memory(alias_idx);
4041 init_in_merge->set_memory_at(alias_idx, prevmem);
4042 if (init_out_raw != nullptr) {
4043 kit.set_memory(init_out_raw, alias_idx);
4044 }
4045 }
4046
4047 //---------------------------set_output_for_allocation-------------------------
4048 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4049 const TypeOopPtr* oop_type,
4050 bool deoptimize_on_exception) {
4051 int rawidx = Compile::AliasIdxRaw;
4052 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4053 add_safepoint_edges(alloc);
4054 Node* allocx = _gvn.transform(alloc);
4055 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4056 // create memory projection for i_o
4057 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4058 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4059
4060 // create a memory projection as for the normal control path
4061 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4062 set_memory(malloc, rawidx);
4063
4064 // a normal slow-call doesn't change i_o, but an allocation does
4065 // we create a separate i_o projection for the normal control path
4066 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4067 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4068
4069 // put in an initialization barrier
4070 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4071 rawoop)->as_Initialize();
4072 assert(alloc->initialization() == init, "2-way macro link must work");
4073 assert(init ->allocation() == alloc, "2-way macro link must work");
4074 {
4075 // Extract memory strands which may participate in the new object's
4076 // initialization, and source them from the new InitializeNode.
4077 // This will allow us to observe initializations when they occur,
4078 // and link them properly (as a group) to the InitializeNode.
4079 assert(init->in(InitializeNode::Memory) == malloc, "");
4080 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4081 init->set_req(InitializeNode::Memory, minit_in);
4082 record_for_igvn(minit_in); // fold it up later, if possible
4083 _gvn.set_type(minit_in, Type::MEMORY);
4084 Node* minit_out = memory(rawidx);
4085 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4086 // Add an edge in the MergeMem for the header fields so an access
4087 // to one of those has correct memory state
4088 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
4089 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
4090 if (oop_type->isa_aryptr()) {
4091 const TypeAryPtr* arytype = oop_type->is_aryptr();
4092 if (arytype->is_flat()) {
4093 // Initially all flat array accesses share a single slice
4094 // but that changes after parsing. Prepare the memory graph so
4095 // it can optimize flat array accesses properly once they
4096 // don't share a single slice.
4097 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4098 C->set_flat_accesses_share_alias(false);
4099 ciInlineKlass* vk = arytype->elem()->inline_klass();
4100 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
4101 ciField* field = vk->nonstatic_field_at(i);
4102 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4103 continue; // do not bother to track really large numbers of fields
4104 int off_in_vt = field->offset_in_bytes() - vk->first_field_offset();
4105 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
4106 int fieldidx = C->get_alias_index(adr_type, true);
4107 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node
4108 // can result in per flat array field Phis to be created which confuses the logic of
4109 // Compile::adjust_flat_array_access_aliases().
4110 hook_memory_on_init(*this, fieldidx, minit_in, nullptr);
4111 }
4112 C->set_flat_accesses_share_alias(true);
4113 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
4114 } else {
4115 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4116 int elemidx = C->get_alias_index(telemref);
4117 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
4118 }
4119 } else if (oop_type->isa_instptr()) {
4120 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
4121 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4122 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4123 ciField* field = ik->nonstatic_field_at(i);
4124 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4125 continue; // do not bother to track really large numbers of fields
4126 // Find (or create) the alias category for this field:
4127 int fieldidx = C->alias_type(field)->index();
4128 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
4129 }
4130 }
4131 }
4132
4133 // Cast raw oop to the real thing...
4134 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4135 javaoop = _gvn.transform(javaoop);
4136 C->set_recent_alloc(control(), javaoop);
4137 assert(just_allocated_object(control()) == javaoop, "just allocated");
4138
4139 #ifdef ASSERT
4140 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4151 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4152 }
4153 }
4154 #endif //ASSERT
4155
4156 return javaoop;
4157 }
4158
4159 //---------------------------new_instance--------------------------------------
4160 // This routine takes a klass_node which may be constant (for a static type)
4161 // or may be non-constant (for reflective code). It will work equally well
4162 // for either, and the graph will fold nicely if the optimizer later reduces
4163 // the type to a constant.
4164 // The optional arguments are for specialized use by intrinsics:
4165 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4166 // - If 'return_size_val', report the total object size to the caller.
4167 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4168 Node* GraphKit::new_instance(Node* klass_node,
4169 Node* extra_slow_test,
4170 Node* *return_size_val,
4171 bool deoptimize_on_exception,
4172 InlineTypeNode* inline_type_node) {
4173 // Compute size in doublewords
4174 // The size is always an integral number of doublewords, represented
4175 // as a positive bytewise size stored in the klass's layout_helper.
4176 // The layout_helper also encodes (in a low bit) the need for a slow path.
4177 jint layout_con = Klass::_lh_neutral_value;
4178 Node* layout_val = get_layout_helper(klass_node, layout_con);
4179 bool layout_is_con = (layout_val == nullptr);
4180
4181 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4182 // Generate the initial go-slow test. It's either ALWAYS (return a
4183 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4184 // case) a computed value derived from the layout_helper.
4185 Node* initial_slow_test = nullptr;
4186 if (layout_is_con) {
4187 assert(!StressReflectiveCode, "stress mode does not use these paths");
4188 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4189 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4190 } else { // reflective case
4191 // This reflective path is used by Unsafe.allocateInstance.
4192 // (It may be stress-tested by specifying StressReflectiveCode.)
4193 // Basically, we want to get into the VM is there's an illegal argument.
4194 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4195 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4196 if (extra_slow_test != intcon(0)) {
4197 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4198 }
4199 // (Macro-expander will further convert this to a Bool, if necessary.)
4210
4211 // Clear the low bits to extract layout_helper_size_in_bytes:
4212 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4213 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4214 size = _gvn.transform( new AndXNode(size, mask) );
4215 }
4216 if (return_size_val != nullptr) {
4217 (*return_size_val) = size;
4218 }
4219
4220 // This is a precise notnull oop of the klass.
4221 // (Actually, it need not be precise if this is a reflective allocation.)
4222 // It's what we cast the result to.
4223 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4224 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4225 const TypeOopPtr* oop_type = tklass->as_instance_type();
4226
4227 // Now generate allocation code
4228
4229 // The entire memory state is needed for slow path of the allocation
4230 // since GC and deoptimization can happen.
4231 Node *mem = reset_memory();
4232 set_all_memory(mem); // Create new memory state
4233
4234 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4235 control(), mem, i_o(),
4236 size, klass_node,
4237 initial_slow_test, inline_type_node);
4238
4239 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4240 }
4241
4242 //-------------------------------new_array-------------------------------------
4243 // helper for newarray and anewarray
4244 // The 'length' parameter is (obviously) the length of the array.
4245 // The optional arguments are for specialized use by intrinsics:
4246 // - If 'return_size_val', report the non-padded array size (sum of header size
4247 // and array body) to the caller.
4248 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4249 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4250 Node* length, // number of array elements
4251 int nargs, // number of arguments to push back for uncommon trap
4252 Node* *return_size_val,
4253 bool deoptimize_on_exception) {
4254 jint layout_con = Klass::_lh_neutral_value;
4255 Node* layout_val = get_layout_helper(klass_node, layout_con);
4256 bool layout_is_con = (layout_val == nullptr);
4257
4258 if (!layout_is_con && !StressReflectiveCode &&
4259 !too_many_traps(Deoptimization::Reason_class_check)) {
4260 // This is a reflective array creation site.
4261 // Optimistically assume that it is a subtype of Object[],
4262 // so that we can fold up all the address arithmetic.
4263 layout_con = Klass::array_layout_helper(T_OBJECT);
4264 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4265 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4266 { BuildCutout unless(this, bol_lh, PROB_MAX);
4267 inc_sp(nargs);
4268 uncommon_trap(Deoptimization::Reason_class_check,
4269 Deoptimization::Action_maybe_recompile);
4270 }
4271 layout_val = nullptr;
4272 layout_is_con = true;
4273 }
4274
4275 // Generate the initial go-slow test. Make sure we do not overflow
4276 // if length is huge (near 2Gig) or negative! We do not need
4277 // exact double-words here, just a close approximation of needed
4278 // double-words. We can't add any offset or rounding bits, lest we
4279 // take a size -1 of bytes and make it positive. Use an unsigned
4280 // compare, so negative sizes look hugely positive.
4281 int fast_size_limit = FastAllocateSizeLimit;
4282 if (layout_is_con) {
4283 assert(!StressReflectiveCode, "stress mode does not use these paths");
4284 // Increase the size limit if we have exact knowledge of array type.
4285 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4286 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4287 }
4288
4289 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4290 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4291
4292 // --- Size Computation ---
4293 // array_size = round_to_heap(array_header + (length << elem_shift));
4294 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4295 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4296 // The rounding mask is strength-reduced, if possible.
4297 int round_mask = MinObjAlignmentInBytes - 1;
4298 Node* header_size = nullptr;
4299 // (T_BYTE has the weakest alignment and size restrictions...)
4300 if (layout_is_con) {
4301 int hsize = Klass::layout_helper_header_size(layout_con);
4302 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4303 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4304 if ((round_mask & ~right_n_bits(eshift)) == 0)
4305 round_mask = 0; // strength-reduce it if it goes away completely
4306 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4307 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4308 assert(header_size_min <= hsize, "generic minimum is smallest");
4309 header_size = intcon(hsize);
4310 } else {
4311 Node* hss = intcon(Klass::_lh_header_size_shift);
4312 Node* hsm = intcon(Klass::_lh_header_size_mask);
4313 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4314 header_size = _gvn.transform(new AndINode(header_size, hsm));
4315 }
4316
4317 Node* elem_shift = nullptr;
4318 if (layout_is_con) {
4319 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4320 if (eshift != 0)
4321 elem_shift = intcon(eshift);
4322 } else {
4323 // There is no need to mask or shift this value.
4324 // The semantics of LShiftINode include an implicit mask to 0x1F.
4325 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4326 elem_shift = layout_val;
4373 }
4374 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4375
4376 if (return_size_val != nullptr) {
4377 // This is the size
4378 (*return_size_val) = non_rounded_size;
4379 }
4380
4381 Node* size = non_rounded_size;
4382 if (round_mask != 0) {
4383 Node* mask1 = MakeConX(round_mask);
4384 size = _gvn.transform(new AddXNode(size, mask1));
4385 Node* mask2 = MakeConX(~round_mask);
4386 size = _gvn.transform(new AndXNode(size, mask2));
4387 }
4388 // else if round_mask == 0, the size computation is self-rounding
4389
4390 // Now generate allocation code
4391
4392 // The entire memory state is needed for slow path of the allocation
4393 // since GC and deoptimization can happen.
4394 Node *mem = reset_memory();
4395 set_all_memory(mem); // Create new memory state
4396
4397 if (initial_slow_test->is_Bool()) {
4398 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4399 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4400 }
4401
4402 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4403 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4404 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4405
4406 // Inline type array variants:
4407 // - null-ok: ciObjArrayKlass with is_elem_null_free() = false
4408 // - null-free: ciObjArrayKlass with is_elem_null_free() = true
4409 // - null-free, flat: ciFlatArrayKlass with is_elem_null_free() = true
4410 // Check if array is a null-free, non-flat inline type array
4411 // that needs to be initialized with the default inline type.
4412 Node* default_value = nullptr;
4413 Node* raw_default_value = nullptr;
4414 if (ary_ptr != nullptr && ary_ptr->klass_is_exact()) {
4415 // Array type is known
4416 if (ary_ptr->is_null_free() && !ary_ptr->is_flat()) {
4417 ciInlineKlass* vk = ary_ptr->elem()->inline_klass();
4418 default_value = InlineTypeNode::default_oop(gvn(), vk);
4419 if (UseCompressedOops) {
4420 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4421 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4422 Node* lower = _gvn.transform(new CastP2XNode(control(), default_value));
4423 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4424 raw_default_value = _gvn.transform(new OrLNode(lower, upper));
4425 } else {
4426 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4427 }
4428 }
4429 }
4430
4431 Node* valid_length_test = _gvn.intcon(1);
4432 if (ary_type->isa_aryptr()) {
4433 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4434 jint max = TypeAryPtr::max_array_length(bt);
4435 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4436 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4437 }
4438
4439 // Create the AllocateArrayNode and its result projections
4440 AllocateArrayNode* alloc
4441 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4442 control(), mem, i_o(),
4443 size, klass_node,
4444 initial_slow_test,
4445 length, valid_length_test,
4446 default_value, raw_default_value);
4447 // Cast to correct type. Note that the klass_node may be constant or not,
4448 // and in the latter case the actual array type will be inexact also.
4449 // (This happens via a non-constant argument to inline_native_newArray.)
4450 // In any case, the value of klass_node provides the desired array type.
4451 const TypeInt* length_type = _gvn.find_int_type(length);
4452 if (ary_type->isa_aryptr() && length_type != nullptr) {
4453 // Try to get a better type than POS for the size
4454 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4455 }
4456
4457 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4458
4459 array_ideal_length(alloc, ary_type, true);
4460 return javaoop;
4461 }
4462
4463 // The following "Ideal_foo" functions are placed here because they recognize
4464 // the graph shapes created by the functions immediately above.
4465
4466 //---------------------------Ideal_allocation----------------------------------
4573 set_all_memory(ideal.merged_memory());
4574 set_i_o(ideal.i_o());
4575 set_control(ideal.ctrl());
4576 }
4577
4578 void GraphKit::final_sync(IdealKit& ideal) {
4579 // Final sync IdealKit and graphKit.
4580 sync_kit(ideal);
4581 }
4582
4583 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4584 Node* len = load_array_length(load_String_value(str, set_ctrl));
4585 Node* coder = load_String_coder(str, set_ctrl);
4586 // Divide length by 2 if coder is UTF16
4587 return _gvn.transform(new RShiftINode(len, coder));
4588 }
4589
4590 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4591 int value_offset = java_lang_String::value_offset();
4592 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4593 false, nullptr, Type::Offset(0));
4594 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4595 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4596 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true),
4597 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4598 Node* p = basic_plus_adr(str, str, value_offset);
4599 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4600 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4601 return load;
4602 }
4603
4604 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4605 if (!CompactStrings) {
4606 return intcon(java_lang_String::CODER_UTF16);
4607 }
4608 int coder_offset = java_lang_String::coder_offset();
4609 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4610 false, nullptr, Type::Offset(0));
4611 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4612
4613 Node* p = basic_plus_adr(str, str, coder_offset);
4614 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4615 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4616 return load;
4617 }
4618
4619 void GraphKit::store_String_value(Node* str, Node* value) {
4620 int value_offset = java_lang_String::value_offset();
4621 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4622 false, nullptr, Type::Offset(0));
4623 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4624
4625 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4626 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4627 }
4628
4629 void GraphKit::store_String_coder(Node* str, Node* value) {
4630 int coder_offset = java_lang_String::coder_offset();
4631 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4632 false, nullptr, Type::Offset(0));
4633 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4634
4635 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4636 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4637 }
4638
4639 // Capture src and dst memory state with a MergeMemNode
4640 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4641 if (src_type == dst_type) {
4642 // Types are equal, we don't need a MergeMemNode
4643 return memory(src_type);
4644 }
4645 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4646 record_for_igvn(merge); // fold it up later, if possible
4647 int src_idx = C->get_alias_index(src_type);
4648 int dst_idx = C->get_alias_index(dst_type);
4649 merge->set_memory_at(src_idx, memory(src_idx));
4650 merge->set_memory_at(dst_idx, memory(dst_idx));
4651 return merge;
4652 }
4725 i_char->init_req(2, AddI(i_char, intcon(2)));
4726
4727 set_control(IfFalse(iff));
4728 set_memory(st, TypeAryPtr::BYTES);
4729 }
4730
4731 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4732 if (!field->is_constant()) {
4733 return nullptr; // Field not marked as constant.
4734 }
4735 ciInstance* holder = nullptr;
4736 if (!field->is_static()) {
4737 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4738 if (const_oop != nullptr && const_oop->is_instance()) {
4739 holder = const_oop->as_instance();
4740 }
4741 }
4742 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4743 /*is_unsigned_load=*/false);
4744 if (con_type != nullptr) {
4745 Node* con = makecon(con_type);
4746 if (field->type()->is_inlinetype()) {
4747 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free());
4748 } else if (con_type->is_inlinetypeptr()) {
4749 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free());
4750 }
4751 return con;
4752 }
4753 return nullptr;
4754 }
4755
4756 //---------------------------load_mirror_from_klass----------------------------
4757 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4758 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4759 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4760 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4761 // mirror = ((OopHandle)mirror)->resolve();
4762 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4763 }
4764
4765 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4766 const Type* obj_type = obj->bottom_type();
4767 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4768 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4769 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4770 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4771 obj = casted_obj;
4772 }
4773 if (sig_type->is_inlinetypeptr()) {
4774 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass(), !gvn().type(obj)->maybe_null());
4775 }
4776 return obj;
4777 }
|